The 193-work starting collection curated by Alfonso Fuggetta, Damian Andrew Tamburri and Alessandro Margara at Software Engineering @ PoliMi. The atlas adds its own summaries, evidence qualifications and graph mappings.
No matching records. Clear the search or choose another group.
W001PolimiIn the published NDSE2 bibliography
Agarwal · 2026
First-adopter agent use raises commits and code volume, but complexity and warnings rise too.
Evidence, context & provenance
Evidence boundaryObservational assignment; output is not delivered value.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Technical debt and complexity · Evaluation methodology and baselines · Long-running agent execution
16 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI IDEs or Autonomous Agents? Measuring the Impact of Coding Agents on Software Development",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Agarwal",
"source": "https://arxiv.org/abs/2601.13597",
"contribution": "First-adopter agent use raises commits and code volume, but complexity and warnings rise too.",
"limitation": "Observational assignment; output is not delivered value.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_TABLES_LOCATED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass2_source_locator": "https://arxiv.org/html/2601.13597",
"pass2_source_version": "arXiv v2 2026-01-27 and published DOI",
"pass2_inspection_scope": "Sections 3.1--3.3 and Table 2 and design limitations",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NO",
"pass2_research_design": "Staggered DiD matched observables using AIDev v3",
"pass2_critical_limitation": "No randomized adoption; selective repositories and adoption attribution; warning IDE-first effect not statistically significant",
"pass2_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C02\", \"C:C05\", \"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W002PolimiIn the published NDSE2 bibliography
Chen · 2026
Delegating to an agent enables tasks beyond copilot assistance; lack of visibility constrains trust.
Evidence, context & provenance
Evidence boundaryTwenty students; completion-time effects unspecified.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Automation bias and trust · Specifications and governance · Skills formation and teaching
31 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Code with Me or for Me? How Increasing AI Automation Transforms Developer Workflows",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Chen",
"source": "https://arxiv.org/abs/2507.08149",
"contribution": "Delegating to an agent enables tasks beyond copilot assistance; lack of visibility constrains trust.",
"limitation": "Twenty students; completion-time effects unspecified.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH04:ABSTRACT",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C03\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W003PolimiIn the published NDSE2 bibliography
Cui · 2026
Real-world coding assistants improve task completion, with an estimated 26% increase across thousands of developers.
Evidence, context & provenance
Evidence boundaryFirst-generation tools; tasks completed do not measure shipped value.
Source type, as recorded: Controlled human or field experiment
Editorial use. Measure activity and delivery separately; require downstream outcomes.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Developer workflow and adoption · Skills formation and teaching · Evaluation methodology and baselines
21 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Cui",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Real-world coding assistants improve task completion, with an estimated 26% increase across thousands of developers.",
"limitation": "First-generation tools; tasks completed do not measure shipped value.",
"manual_use": "Measure activity and delivery separately; require downstream outcomes.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass3_audit_primary_url": "https://pubsonline.informs.org/doi/10.1287/mnsc.2025.00535",
"pass3_audit_version": "Management Science 2026; author-hosted working paper June 2025 Table 3",
"pass3_audit_depth": "WORKING_PAPER_TABLE_SELECTED_PUBLISHER_ABSTRACT",
"pass3_audit_methods_locators": "2025 author manuscript Table 3 and publisher abstract",
"pass3_audit_sample": "Three company randomized rollout designs, 4,867 engineers",
"pass3_audit_observation": "Company WIV estimates 27.38,17.94,54.03%; pooled 26.08% SE10.3",
"pass3_audit_boundary": "Outcome in working Table 3 is pull requests; journal full table parity outstanding",
"pass3_audit_original_data_replication": "NO",
"pass3_audit_provenance": "PASS3_PRINTED_EDITION_RECONSTRUCTED_2026-10-10",
"pass3_audit_date": "2026-10-10",
"pass9_scope": "publisher official 15-page online supplement pp24-25,28; sample+ITT checked; no regressions refit",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C02\", \"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W004PolimiIn the published NDSE2 bibliography
He · 2026
Early code-volume surge disappears as complexity and static-analysis findings remain elevated.
Evidence, context & provenance
Evidence boundaryMatched OSS repositories cannot remove all adoption confounding.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Measure activity and delivery separately; require downstream outcomes.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Developer workflow and adoption · Technical debt and complexity · Economic costs and return · Evaluation methodology and baselines
20 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Speed at the Cost of Quality: How Cursor AI Increases Short-Term Velocity and Long-Term Complexity in Open-Source Projects",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "He",
"source": "https://arxiv.org/abs/2511.04427",
"contribution": "Early code-volume surge disappears as complexity and static-analysis findings remain elevated.",
"limitation": "Matched OSS repositories cannot remove all adoption confounding.",
"manual_use": "Measure activity and delivery separately; require downstream outcomes.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:TARGETED_FULL_TEXT_PAGES_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass2_source_locator": "https://www.cs.cmu.edu/~ckaestne/pdf/msr26.pdf",
"pass2_source_version": "MSR 2026 publisher DOI and 19-page author PDF",
"pass2_inspection_scope": "Sections 3.5 and 4.1 Table 2 printed page 7; Appendix B.3 Table 6 printed page 16",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NO",
"pass2_research_design": "Matched staggered DiD with three nonidentical estimators",
"pass2_critical_limitation": "Estimator choices materially change quality findings; config-file use not exact agent activity; nonrandom treatment",
"pass2_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C02\", \"C:C05\", \"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W005PolimiIn the published NDSE2 bibliography
Raida · 2026
Real pull-request usage is concentrated and frequently consists of just one or two agent PRs per adopting project.
Evidence, context & provenance
Evidence boundaryEarly adoption snapshot; does not measure organizational productivity.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Developer workflow and adoption · Evaluation methodology and baselines
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Early Adoption of Agentic Coding Tools by GitHub Projects",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Raida",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Real pull-request usage is concentrated and frequently consists of just one or two agent PRs per adopting project.",
"limitation": "Early adoption snapshot; does not measure organizational productivity.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W006PolimiIn the published NDSE2 bibliography
Robbes · 2026
Coding-agent participation is spreading across public repositories and changes commit size and activity patterns.
Evidence, context & provenance
Evidence boundaryAdoption and size are not effectiveness.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Agentic Much? Adoption of Coding Agents on GitHub",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Robbes",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Coding-agent participation is spreading across public repositories and changes commit size and activity patterns.",
"limitation": "Adoption and size are not effectiveness.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W007PolimiIn the published NDSE2 bibliography
Merler · 2026
Enterprise users report perceived usefulness and practical requirements that should shape deployment.
Evidence, context & provenance
Evidence boundaryIBM respondents only; perception-based.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Measure activity and delivery separately; require downstream outcomes.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Automation bias and trust
23 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Usage, Effects and Requirements for AI Coding Assistants in the Enterprise: An Empirical Study",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Merler",
"source": "https://arxiv.org/abs/2601.20112",
"contribution": "Enterprise users report perceived usefulness and practical requirements that should shape deployment.",
"limitation": "IBM respondents only; perception-based.",
"manual_use": "Measure activity and delivery separately; require downstream outcomes.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://arxiv.org/abs/2601.20112",
"pass7_identity_source_class": "author_preprint",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W008PolimiIn the published NDSE2 bibliography
Paradis · 2025
Controlled enterprise tasks indicate roughly 21% faster task performance with assistance.
Evidence, context & provenance
Evidence boundaryNarrow tasks and 96 participants limit generalization to long-lived systems.
Source type, as recorded: Controlled human or field experiment
Editorial use. Measure activity and delivery separately; require downstream outcomes.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "How Much Does AI Impact Development Speed? An Enterprise-Based Randomized Controlled Trial",
"kind": "paper",
"group": "Productivity",
"year": "2025",
"lead_author": "Paradis",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Controlled enterprise tasks indicate roughly 21% faster task performance with assistance.",
"limitation": "Narrow tasks and 96 participants limit generalization to long-lived systems.",
"manual_use": "Measure activity and delivery separately; require downstream outcomes.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2025",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C02\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W009PolimiIn the published NDSE2 bibliography
Ziegler · 2022
Suggestion acceptance is only weakly associated with self-rated productivity, challenging acceptance rate as KPI.
Evidence, context & provenance
Evidence boundaryVendor users; survey not causal.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Productivity Assessment of Neural Code Completion",
"kind": "paper",
"group": "Productivity",
"year": "2022",
"lead_author": "Ziegler",
"source": "https://arxiv.org/abs/2205.06537",
"contribution": "Suggestion acceptance is only weakly associated with self-rated productivity, challenging acceptance rate as KPI.",
"limitation": "Vendor users; survey not causal.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2022",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C02\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W010PolimiIn the published NDSE2 bibliography
Demirer · 2026
Agent advances raise commits substantially more than releases; delivery throughput has lower elasticity than coding throughput.
Evidence, context & provenance
Evidence boundaryGitHub activity and matched designs do not establish business value.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Evaluation methodology and baselines
38 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Demirer",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Agent advances raise commits substantially more than releases; delivery throughput has lower elasticity than coding throughput.",
"limitation": "GitHub activity and matched designs do not establish business value.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no_institutional_record_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass2_source_locator": "https://www.nber.org/papers/w35275",
"pass2_source_version": "NBER Working Paper 35275 September 2026 revision institutional record",
"pass2_inspection_scope": "Institutional metadata and abstract; complete Sept manuscript not independently inspected",
"pass2_methods_status": "PRIMARY_INSTITUTIONAL_RECORD_ONLY",
"pass2_replication_status": "NO",
"pass2_research_design": "Matched event study plus economic structural model as described by institution",
"pass2_critical_limitation": "Observational adoption and different output endpoints; full revision and exact attribution tables not examined",
"pass2_audit_date": "2026-10-10",
"pass5_audit_date": "2026-10-10",
"pass5_version": "NBER w35275 Sept 2026 revision; SSRN May 2026 separate abstract",
"pass5_primary_locator": "https://www.nber.org/papers/w35275 ; https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6843068",
"pass5_inspection_depth": "PUBLISHER_REVISION_ABSTRACT_AND_EARLIER_SSRN_ABSTRACT_ONLY",
"pass5_page_scope": "Institutional abstracts only; 134-page September PDF not accessible",
"pass5_sample": "September >500,000 developers; earlier May >100,000; distinct samples/estimates",
"pass5_main_finding": "September cumulative commits ~30/180/240%; agent projects +80%; agent releases +30%; elasticity 0.23. May 40/140/180%; projects +50%; releases +30%; elasticity 0.25.",
"pass5_residual_limitation": "No original coefficient tables or microdata inspected; estimates observational and structural, not direct firm outcomes.",
"pass5_original_data_replicated": "NO",
"pass5_provenance": "PASS5_2026-10-10_SCOPED_PRIMARY_AND_VERSIONED_ABSTRACT",
"pass6_date": "2026-10-10",
"pass6_locator": "https://www.nber.org/papers/w35275",
"pass6_inspection_depth": "NBER_PRIMARY_INSTITUTIONAL_REVISED_RECORD",
"pass6_page_scope": "NBER September 2026 revision abstract and author disclosures",
"pass6_finding": "500k plus developer sample; 30/180/240% commit estimates; declared paid Microsoft consulting for two authors",
"pass6_limitation": "SEPT_134PAGE_PDF_UNAVAILABLE",
"pass6_original_data_replicated": "NO",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C02\", \"C:C08\", \"C:C11\", \"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W011PolimiIn the published NDSE2 bibliography
Farrag · 2026
Proposes specification and verification governance to balance faster generation against reliability.
Evidence, context & provenance
Evidence boundarySynthesis rather than a new intervention trial.
Source type, as recorded: Multivocal synthesis + illustrative uncontrolled pilot
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Independent verification and oracles · Specifications and governance · Skills formation and teaching
31 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Productivity-Reliability Paradox: Specification-Driven Governance for AI-Augmented Software Development",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Farrag",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Proposes specification and verification governance to balance faster generation against reliability.",
"limitation": "Synthesis rather than a new intervention trial.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Multivocal synthesis + illustrative uncontrolled pilot",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass2_source_locator": "https://arxiv.org/html/2605.01160",
"pass2_source_version": "arXiv 2605.01160 HTML primary",
"pass2_inspection_scope": "Methods sections 3 and evaluation sections 7.4.1--7.4.5",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NO",
"pass2_research_design": "Multivocal literature review plus nonrandom illustrative before-after exercise",
"pass2_critical_limitation": "Self-selected synthetic/illustrative case with no randomized control and no credible causal estimate",
"pass2_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C03\", \"C:C04\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W012PolimiIn the published NDSE2 bibliography
Liu · 2026
Millions of traces characterize tool usage, agent sessions and infrastructure optimizations.
Evidence, context & provenance
Evidence boundarySession telemetry does not establish downstream outcomes.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Long-running agent execution · Dataset provenance and dependency
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Agentic Coding in the Wild: Characterizing GitHub Copilot Traces at Production Scale",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Liu",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Millions of traces characterize tool usage, agent sessions and infrastructure optimizations.",
"limitation": "Session telemetry does not establish downstream outcomes.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C03\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W013PolimiIn the published NDSE2 bibliography
Lobentanzer · 2026
Compares agent expectations with measured productivity and warns against extrapolating benchmark success.
Evidence, context & provenance
Evidence boundaryReinterpretation of existing measures, not new experiment.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Benchmarks and construct validity · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Quantifying the Expectation-Realisation Gap for Agentic AI Systems",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Lobentanzer",
"source": "https://arxiv.org/abs/2602.20292",
"contribution": "Compares agent expectations with measured productivity and warns against extrapolating benchmark success.",
"limitation": "Reinterpretation of existing measures, not new experiment.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/abs/2602.20292",
"publication_status_note": "arXiv abstract, not full paper audited",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PREPRINT_PRIMARY_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W014PolimiIn the published NDSE2 bibliography
Murphy-Hill · 2026
CLI agent adoption associates with 24% more merged PRs per engineer without detectable four-month decay.
Evidence, context & provenance
Evidence boundaryMerged PRs are a proxy; adopter and tenure selection remain.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Developer workflow and adoption · Benchmarks and construct validity · Evaluation methodology and baselines
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Murphy-Hill",
"source": "https://arxiv.org/abs/2607.01418",
"contribution": "CLI agent adoption associates with 24% more merged PRs per engineer without detectable four-month decay.",
"limitation": "Merged PRs are a proxy; adopter and tenure selection remain.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W015PolimiIn the published NDSE2 bibliography
Quispe · 2026
Claude Code adoption correlates with developers contributing in about 2.5 additional languages; breadth is not mastery.
Evidence, context & provenance
Evidence boundaryRepository activity cannot prove actual acquired proficiency.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Skills formation and teaching
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Agentic Delegation and the Language Frontier of Software Developers: A Model and Evidence from Claude Code on GitHub",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Quispe",
"source": "https://arxiv.org/abs/2605.25438",
"contribution": "Claude Code adoption correlates with developers contributing in about 2.5 additional languages; breadth is not mastery.",
"limitation": "Repository activity cannot prove actual acquired proficiency.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://arxiv.org/abs/2605.25438",
"pass7_identity_source_class": "author_preprint",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W016PolimiIn the published NDSE2 bibliography
Vella · 2026
Most developers report less writing and greater productivity; negative experience measures also increase over time.
Evidence, context & provenance
Evidence boundary95 respondents and self-report.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Developer workflow and adoption · Evaluation methodology and baselines
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Impact of AI Coding Assistants on Software Engineering: A Longitudinal Study",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Vella",
"source": "https://arxiv.org/abs/2605.23135",
"contribution": "Most developers report less writing and greater productivity; negative experience measures also increase over time.",
"limitation": "95 respondents and self-report.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W017PolimiIn the published NDSE2 bibliography
Bono · 2025
Agent support improves administrative throughput and accuracy in enterprise configuration work.
Evidence, context & provenance
Evidence boundaryDifferent domain from programming; do not present as SWE causal evidence.
Source type, as recorded: Controlled human or field experiment
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Randomized Controlled Trials for Conditional Access Optimization Agent",
"kind": "paper",
"group": "Productivity",
"year": "2025",
"lead_author": "Bono",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Agent support improves administrative throughput and accuracy in enterprise configuration work.",
"limitation": "Different domain from programming; do not present as SWE causal evidence.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2025",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "Identity-administration agent RCT (not developer productivity)",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W018PolimiIn the published NDSE2 bibliography
Kumar · 2025
Reports increased engineering output and shorter review cycles under AI assistance.
Evidence, context & provenance
Evidence boundaryUncontrolled before/after changes; difficult attribution.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Review capacity and maintainers · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Intuition to Evidence: Measuring AI's True Impact on Developer Productivity",
"kind": "paper",
"group": "Productivity",
"year": "2025",
"lead_author": "Kumar",
"source": "https://arxiv.org/abs/2509.19708",
"contribution": "Reports increased engineering output and shorter review cycles under AI assistance.",
"limitation": "Uncontrolled before/after changes; difficult attribution.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://arxiv.org/abs/2509.19708",
"pass7_identity_source_class": "author_preprint",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W019PolimiIn the published NDSE2 bibliography
Gambacorta · 2024
Code-generation tool increases output notably for junior developers, suggesting heterogeneous gains.
Evidence, context & provenance
Evidence boundaryOutput-count metric; model and organization specificity.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Employment and early-career roles
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Generative AI and labour productivity: a field experiment on coding",
"kind": "paper",
"group": "Productivity",
"year": "2024",
"lead_author": "Gambacorta",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Code-generation tool increases output notably for junior developers, suggesting heterogeneous gains.",
"limitation": "Output-count metric; model and organization specificity.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2024",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass6_primary_locator": "https://www.bis.org/publications/working-paper-1208-generative-ai-and-labour-productivity-field-experiment-coding",
"pass6_metadata_depth": "PRIMARY_WORKING_PAPER_AND_LATER_JOURNAL",
"pass6_supporting_locator": "https://www.sciencedirect.com/science/article/pii/S1572308926000458",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W020PolimiIn the published NDSE2 bibliography
Becker · 2026
Survey respondents report major time savings with substantial heterogeneity.
Evidence, context & provenance
Evidence boundaryRetrospective self estimates are not measured time saved.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Evaluation methodology and baselines
17 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Becker",
"source": "https://metr.org/blog/2026-05-11-ai-usage-survey/",
"contribution": "Survey respondents report major time savings with substantial heterogeneity.",
"limitation": "Retrospective self estimates are not measured time saved.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://metr.org/blog/2026-05-11-ai-usage-survey/",
"publication_status_note": "METR survey, self-reported impact; not causal or timed productivity",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_INSTITUTION_REPORT",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"pass5_audit_date": "2026-10-10",
"pass5_version": "METR May 11 2026 original research blog and appendix",
"pass5_primary_locator": "https://metr.substack.com/p/2026-05-11-ai-usage-survey",
"pass5_inspection_depth": "PRIMARY_METHODS_SURVEY_QUESTION_WORDING_FILTERING_AND_RESULTS",
"pass5_page_scope": "HTML Summary, Methodology, Results, Data quality and footnotes",
"pass5_sample": "359 raw unique respondents, 10 excluded anomalous -> 349 retained; convenience recruitment; approx 2% emailed response; approx 70% compensated",
"pass5_main_finding": "Median self-reported value multiplier 1.4–2x and speed multiplier ~3x; retrospective 2025 1.3x, 2027 prediction 2.5x; these are beliefs and cannot be compared as experimental gains.",
"pass5_residual_limitation": "No independent observation-based validation, raw respondent-level data not rerun; four listed occupational counts total 335 not the full retained 349, without proof of discrepancy.",
"pass5_original_data_replicated": "NO",
"pass5_provenance": "PASS5_2026-10-10_SCOPED_PRIMARY_AND_VERSIONED_ABSTRACT",
"ndse2_actual_chapters_json": "[\"C:C02\", \"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W021PolimiIn the published NDSE2 bibliography
Faros AI · 2026
Higher AI use associates with more throughput, larger PRs, more incidents and much slower review.
Evidence, context & provenance
Evidence boundaryClosed vendor data; observational metrics and unreported definitions.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Dataset provenance and dependency
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The AI Engineering Report 2026: The Acceleration Whiplash",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Faros AI",
"source": "https://www.faros.ai/blog/ai-acceleration-whiplash-takeaways",
"contribution": "Higher AI use associates with more throughput, larger PRs, more incidents and much slower review.",
"limitation": "Closed vendor data; observational metrics and unreported definitions.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://www.faros.ai/blog/ai-acceleration-whiplash-takeaways",
"pass7_identity_source_class": "producer_primary_report",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W022PolimiIn the published NDSE2 bibliography
JetBrains Research · 2026
Frequent agent use has quickly become mainstream in the surveyed population.
Evidence, context & provenance
Evidence boundaryAdoption levels reflect sample and do not measure value or quality.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Evaluation methodology and baselines · Dataset provenance and dependency
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI Coding Agents: Adoption Trends",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "JetBrains Research",
"source": "https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/",
"contribution": "Frequent agent use has quickly become mainstream in the surveyed population.",
"limitation": "Adoption levels reflect sample and do not measure value or quality.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/",
"pass7_identity_source_class": "producer_primary_survey",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W023PolimiIn the published NDSE2 bibliography
Becker · 2025
Experienced maintainers were approximately 19% slower on familiar real-repository tasks using early-2025 tools.
Evidence, context & provenance
Evidence boundarySmall sample and older tool generation; not universal effect.
Source type, as recorded: Controlled human or field experiment
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Review capacity and maintainers · Evaluation methodology and baselines · Dataset provenance and dependency
26 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity",
"kind": "paper",
"group": "Productivity",
"year": "2025",
"lead_author": "Becker",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Experienced maintainers were approximately 19% slower on familiar real-repository tasks using early-2025 tools.",
"limitation": "Small sample and older tool generation; not universal effect.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2025",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C02\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W024PolimiIn the published NDSE2 bibliography
Google Cloud DORA · 2025
AI use links with greater throughput but lower stability; existing organizational capability moderates outcomes.
Evidence, context & provenance
Evidence boundaryCross-sectional self-report; association not causal.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Long-running agent execution
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "2025 State of AI-Assisted Software Development",
"kind": "paper",
"group": "Productivity",
"year": "2025",
"lead_author": "Google Cloud DORA",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "AI use links with greater throughput but lower stability; existing organizational capability moderates outcomes.",
"limitation": "Cross-sectional self-report; association not causal.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2025",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass6_primary_locator": "https://dora.dev/research/2025/dora-report/",
"pass6_metadata_depth": "OFFICIAL_REPORT_INDEX",
"pass6_supporting_locator": "https://dora.dev/research/2025/errata/",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W025PolimiIn the published NDSE2 bibliography
Tamkin · 2025
Claude conversation analysis suggests sizable theoretical time saving.
Evidence, context & provenance
Evidence boundaryCounterfactual completion time is estimated by model, not observed.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Evaluation methodology and baselines
13 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Estimating AI productivity gains from Claude conversations",
"kind": "paper",
"group": "Productivity",
"year": "2025",
"lead_author": "Tamkin",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "Claude conversation analysis suggests sizable theoretical time saving.",
"limitation": "Counterfactual completion time is estimated by model, not observed.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass6_primary_locator": "https://www.anthropic.com/research/estimating-productivity-gains",
"pass6_metadata_depth": "PRIMARY_RESEARCH_NOTE_AND_PDF",
"pass6_supporting_locator": "https://assets.anthropic.com/m/28fda2ad148e2bf5/original/Estimating-AI-productivity-gains-from-Claude-conversations.pdf",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W026PolimiIn the published NDSE2 bibliography
Google Cloud DORA · 2024
AI uptake associated with modest delivery-throughput and stability deterioration, highlighting systems bottlenecks.
Evidence, context & provenance
Evidence boundaryCorrelational; differs from later DORA populations and tools.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Long-running agent execution
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Accelerate State of DevOps Report 2024",
"kind": "paper",
"group": "Productivity",
"year": "2024",
"lead_author": "Google Cloud DORA",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"contribution": "AI uptake associated with modest delivery-throughput and stability deterioration, highlighting systems bottlenecks.",
"limitation": "Correlational; differs from later DORA populations and tools.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2024",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass6_primary_locator": "https://dora.dev/research/2024/dora-report/",
"pass6_metadata_depth": "OFFICIAL_REPORT_INDEX",
"pass6_supporting_locator": "https://dora.dev/research/2024/errata/",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W027PolimiIn the published NDSE2 bibliography
Relyea · 2026
Some organizations accelerate while a substantial subset reports declining productivity.
Evidence, context & provenance
Evidence boundarySelf-reported by leaders; selection and measurement limitations.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Developer workflow and adoption · Evaluation methodology and baselines
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Beyond the Copilot: Scaling the Agentic Product Development Life Cycle",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Relyea",
"source": "https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/beyond-the-copilot-scaling-the-agentic-product-development-life-cycle",
"contribution": "Some organizations accelerate while a substantial subset reports declining productivity.",
"limitation": "Self-reported by leaders; selection and measurement limitations.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/beyond-the-copilot-scaling-the-agentic-product-development-life-cycle",
"pass7_identity_source_class": "producer_primary_report",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-productivity",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W028PolimiIn the published NDSE2 bibliography
Relyea · 2026
Advocates transforming the development lifecycle rather than optimizing code generation alone.
Evidence, context & provenance
Evidence boundaryExtrapolated and survey-based benefits are not verified causal estimates.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Measure activity and delivery separately; require downstream outcomes.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Independent verification and oracles · Architecture and modularity
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The AI Revolution in Software Development",
"kind": "paper",
"group": "Productivity",
"year": "2026",
"lead_author": "Relyea",
"source": "https://www.mckinsey.com/~/media/mckinsey/business%20functions/tech%20and%20ai/our%20insights/the%20ai%20revolution%20in%20software%20development/the-ai-revolution-in-software-development_final.pdf",
"contribution": "Advocates transforming the development lifecycle rather than optimizing code generation alone.",
"limitation": "Extrapolated and survey-based benefits are not verified causal estimates.",
"manual_use": "Measure activity and delivery separately; require downstream outcomes.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://www.mckinsey.com/~/media/mckinsey/business%20functions/tech%20and%20ai/our%20insights/the%20ai%20revolution%20in%20software%20development/the-ai-revolution-in-software-development_final.pdf",
"publication_status_note": "Industry commentary, not a randomized coding experiment",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_INSTITUTION_REPORT",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass5_audit_date": "2026-10-10",
"pass5_version": "McKinsey Tech & AI Practice April 2026 book excerpt (nine-page original PDF)",
"pass5_primary_locator": "https://www.mckinsey.com/~/media/mckinsey/business%20functions/tech%20and%20ai/our%20insights/the%20ai%20revolution%20in%20software%20development/the-ai-revolution-in-software-development_final.pdf",
"pass5_inspection_depth": "PRIMARY_FULL_NINE_PAGE_PDF_SELECT_METRIC_CONTEXT",
"pass5_page_scope": "PDF pp.1-4, esp. pp.2-4 (vignette, exhibit, organizational analysis)",
"pass5_sample": "Anonymized bank vignette; illustrative four-level leverage exhibit; analysis of nearly 300 public companies cited from separate 2025 work",
"pass5_main_finding": "20x figure describes next-frontier potential; text describes Level 4 as largely experimental; top-quintile 16–30% and 31–45% claim comes from separately cited cross-company material, not the same 20x trial.",
"pass5_residual_limitation": "No linked raw implementation metrics, confidence intervals or independent observational validation in this nine-page excerpt; must not call 20x a measured population mean.",
"pass5_original_data_replicated": "NO",
"pass5_provenance": "PASS5_2026-10-10_SCOPED_PRIMARY_AND_VERSIONED_ABSTRACT",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W029PolimiIn the published NDSE2 bibliography
Cynthia · 2026
Merged agent PRs still introduce code smells and bugs; size of change dominates agent-specific differences.
Evidence, context & provenance
Evidence boundaryStatic-analysis findings are not confirmed operational incidents.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Technical debt and complexity
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Beyond Bug Fixes: An Empirical Investigation of Post-Merge Code Quality Issues in Agent-Generated Pull Requests",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Cynthia",
"source": "https://arxiv.org/abs/2601.20109",
"contribution": "Merged agent PRs still introduce code smells and bugs; size of change dominates agent-specific differences.",
"limitation": "Static-analysis findings are not confirmed operational incidents.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:TARGETED_FULL_TEXT_PAGES_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W030PolimiIn the published NDSE2 bibliography
Day · 2026
Modular workspace layouts reduce indirect injection success, indicating architecture can be a security control.
Evidence, context & provenance
Evidence boundaryOnly open-weight agents; no transfer figures in bibliography.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Vulnerability and secure code · Prompt injection and malicious tools · Architecture and modularity · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Workspace Topology as an Attack Vector in Agentic Coding Assistants",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Day",
"source": "https://arxiv.org/abs/2608.14876",
"contribution": "Modular workspace layouts reduce indirect injection success, indicating architecture can be a security control.",
"limitation": "Only open-weight agents; no transfer figures in bibliography.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W031PolimiIn the published NDSE2 bibliography
Huang · 2026
AI-generated changes miss existing reuse opportunities, creating unnecessary new code.
Evidence, context & provenance
Evidence boundaryNumerical effects and sentiment findings not supplied in bibliography.
Source type, as recorded: Missing outcome details
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Agent failure and escalation · Code reuse and duplication
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "More Code, Less Reuse: Investigating Code Quality and Reviewer Sentiment towards AI-generated Pull Requests",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Huang",
"source": "https://arxiv.org/abs/2601.21276",
"contribution": "AI-generated changes miss existing reuse opportunities, creating unnecessary new code.",
"limitation": "Numerical effects and sentiment findings not supplied in bibliography.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Missing outcome details",
"editorial_priority": "VERIFY",
"bibliography_year": "2026",
"original_evidence_class": "Missing outcome details",
"reviewed_evidence_class": "Missing outcome details",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:TARGETED_FULL_TEXT_PAGES_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W032PolimiIn the published NDSE2 bibliography
Mousavi · 2026
Copilot improves functional correctness without reliably improving secure API use or developer risk recognition.
Evidence, context & provenance
Evidence boundaryForty-four participants and selected security APIs.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Independent verification and oracles · Vulnerability and secure code · System understanding and rationale · Evaluation methodology and baselines
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Understanding the Impact of AI Code Assistants on Security API Usage",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Mousavi",
"source": "https://arxiv.org/abs/2607.11348",
"contribution": "Copilot improves functional correctness without reliably improving secure API use or developer risk recognition.",
"limitation": "Forty-four participants and selected security APIs.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W033PolimiIn the published NDSE2 bibliography
Ogenrwot · 2026
Agent pull requests contain more commits and different change structures than human PRs.
Evidence, context & provenance
Evidence boundaryStructural contrasts are not defect-rate comparisons; shares AIDev cohort.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Architecture and modularity · Evaluation methodology and baselines · Dataset provenance and dependency
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "How AI Coding Agents Modify Code: A Large-Scale Study of GitHub Pull Requests",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Ogenrwot",
"source": "https://arxiv.org/abs/2601.17581",
"contribution": "Agent pull requests contain more commits and different change structures than human PRs.",
"limitation": "Structural contrasts are not defect-rate comparisons; shares AIDev cohort.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_LOCATED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W034PolimiIn the published NDSE2 bibliography
Pinna · 2026
PR acceptance depends markedly on task category; measured agent differences are less robust.
Evidence, context & provenance
Evidence boundaryAcceptance reflects maintainer decisions rather than true quality; shared AIDev data.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Evaluation methodology and baselines · Dataset provenance and dependency
14 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Comparing AI Coding Agents: A Task-Stratified Analysis of Pull Request Acceptance",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Pinna",
"source": "https://arxiv.org/abs/2602.08915",
"contribution": "PR acceptance depends markedly on task category; measured agent differences are less robust.",
"limitation": "Acceptance reflects maintainer decisions rather than true quality; shared AIDev data.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH05:FULL_TEXT_OPENED_METHODS_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W035PolimiIn the published NDSE2 bibliography
Siddiq · 2026
Agent security contributions emphasize hardening, face long reviews, and are affected by PR complexity.
Evidence, context & provenance
Evidence boundaryHardening versus vulnerability fixing must be kept separate; shared AIDev data.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Technical debt and complexity · Vulnerability and secure code · Dataset provenance and dependency
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Security in the Age of AI Teammates: An Empirical Study of Agentic Pull Requests on GitHub",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Siddiq",
"source": "https://arxiv.org/abs/2601.00477",
"contribution": "Agent security contributions emphasize hardening, face long reviews, and are affected by PR complexity.",
"limitation": "Hardening versus vulnerability fixing must be kept separate; shared AIDev data.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W036PolimiIn the published NDSE2 bibliography
Watanabe · 2026
Most Claude Code PRs merge, but nearly half of merged changes undergo human revision.
Evidence, context & provenance
Evidence boundarySmall initial adoption cohort; merge rate is not field reliability.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "On the Use of Agentic Coding: An Empirical Study of Pull Requests on GitHub",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Watanabe",
"source": "https://arxiv.org/abs/2509.14745",
"contribution": "Most Claude Code PRs merge, but nearly half of merged changes undergo human revision.",
"limitation": "Small initial adoption cohort; merge rate is not field reliability.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:TARGETED_FULL_TEXT_PAGES_CHECKED; CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W037PolimiIn the published NDSE2 bibliography
Xiong · 2026
Agent add-on scanners disagree widely and can miss malicious or unsafe skills.
Evidence, context & provenance
Evidence boundarySecurity of extensions is not security of generated code.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Agent failure and escalation · Vulnerability and secure code · Prompt injection and malicious tools · Dependencies and agent supply chain · Specifications and governance · Skills formation and teaching
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "After the Party: Growth, Governance, and Security Scanning in the OpenClaw Agent Skill Ecosystem",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Xiong",
"source": "https://arxiv.org/abs/2609.17274",
"contribution": "Agent add-on scanners disagree widely and can miss malicious or unsafe skills.",
"limitation": "Security of extensions is not security of generated code.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W038PolimiIn the published NDSE2 bibliography
Yan · 2026
External, localized security feedback improves model vulnerability remediation more consistently than self-explanation.
Evidence, context & provenance
Evidence boundaryControlled benchmarks, residual risk and model sensitivity remain.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Vulnerability and secure code · Evaluation methodology and baselines
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Guiding AI to Fix Its Own Flaws: An Empirical Study on LLM-Driven Secure Code Generation",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Yan",
"source": "https://arxiv.org/abs/2506.23034",
"contribution": "External, localized security feedback improves model vulnerability remediation more consistently than self-explanation.",
"limitation": "Controlled benchmarks, residual risk and model sensitivity remain.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:PDF_PAGE_RELEVANT; CH07:FULLTEXT_HTML",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\", \"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W039PolimiIn the published NDSE2 bibliography
Cotroneo · 2025
AI and human code exhibit different fault types: AI more unused constructs and severe vulnerabilities; humans more structural complexity.
Evidence, context & provenance
Evidence boundaryAuthorship and sampling matter; not a universal quality hierarchy.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Convert the reported risk into a static, behavioral or security acceptance gate.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Technical debt and complexity · Vulnerability and secure code · Architecture and modularity · Ownership and knowledge transfer · Dataset provenance and dependency
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Human-Written vs. AI-Generated Code: A Large-Scale Study of Defects, Vulnerabilities, and Complexity",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Cotroneo",
"source": "https://arxiv.org/abs/2508.21634",
"contribution": "AI and human code exhibit different fault types: AI more unused constructs and severe vulnerabilities; humans more structural complexity.",
"limitation": "Authorship and sampling matter; not a universal quality hierarchy.",
"manual_use": "Convert the reported risk into a static, behavioral or security acceptance gate.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:TARGETED_FULL_TEXT_PAGES_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W040PolimiIn the published NDSE2 bibliography
Fu · 2025
Around a quarter to a third of selected AI-attributed snippets trigger security weaknesses; static warnings support partial remediation.
Evidence, context & provenance
Evidence boundaryNo comparable human-authored control.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Vulnerability and secure code · Evaluation methodology and baselines
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Security Weaknesses of Copilot-Generated Code in GitHub Projects: An Empirical Study",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Fu",
"source": "https://arxiv.org/abs/2310.02059",
"contribution": "Around a quarter to a third of selected AI-attributed snippets trigger security weaknesses; static warnings support partial remediation.",
"limitation": "No comparable human-authored control.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W041PolimiIn the published NDSE2 bibliography
Schreiber · 2025
Most selected files have no mapped vulnerability, while thousands of CWE instances remain and rates vary by language.
Evidence, context & provenance
Evidence boundaryNo matched human baseline and marker-based selection.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Vulnerability and secure code · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Security Vulnerabilities in AI-Generated Code: A Large-Scale Analysis of Public GitHub Repositories",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Schreiber",
"source": "https://arxiv.org/abs/2510.26103",
"contribution": "Most selected files have no mapped vulnerability, while thousands of CWE instances remain and rates vary by language.",
"limitation": "No matched human baseline and marker-based selection.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2025",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:PDF_PAGE_RELEVANT",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W042PolimiIn the published NDSE2 bibliography
Shukla · 2025
Iteration intended to improve code can accumulate critical security vulnerabilities if no external gate interrupts it.
Evidence, context & provenance
Evidence boundaryTen crafted starting programs; ecological validity limited.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Vulnerability and secure code · Employment and early-career roles
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Security Degradation in Iterative AI Code Generation: A Systematic Analysis of the Paradox",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Shukla",
"source": "https://arxiv.org/abs/2506.11022",
"contribution": "Iteration intended to improve code can accumulate critical security vulnerabilities if no external gate interrupts it.",
"limitation": "Ten crafted starting programs; ecological validity limited.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W043PolimiIn the published NDSE2 bibliography
Spracklen · 2025
Nonexistent package names create supply-chain takeover possibilities, requiring registry existence and dependency-policy gates.
Evidence, context & provenance
Evidence boundaryModel generation and repository dependency adoption are different endpoints.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Convert the reported risk into a static, behavioral or security acceptance gate.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Benchmarks and construct validity · Dependencies and agent supply chain · Specifications and governance · Human-agent collaboration
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Spracklen",
"source": "https://arxiv.org/abs/2406.10279",
"contribution": "Nonexistent package names create supply-chain takeover possibilities, requiring registry existence and dependency-policy gates.",
"limitation": "Model generation and repository dependency adoption are different endpoints.",
"manual_use": "Convert the reported risk into a static, behavioral or security acceptance gate.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:PDF_PAGE_RELEVANT",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W044PolimiIn the published NDSE2 bibliography
Wu · 2025
Some generated fragments replicate code almost exactly, raising reuse, maintenance and provenance concerns.
Evidence, context & provenance
Evidence boundaryGenerator outputs measured; longitudinal repository clone prevalence unknown.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Convert the reported risk into a static, behavioral or security acceptance gate.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Code reuse and duplication · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "An Empirical Study of Code Clones from Commercial AI Code Generators",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Wu",
"source": "https://doi.org/10.1145/3729397",
"contribution": "Some generated fragments replicate code almost exactly, raising reuse, maintenance and provenance concerns.",
"limitation": "Generator outputs measured; longitudinal repository clone prevalence unknown.",
"manual_use": "Convert the reported risk into a static, behavioral or security acceptance gate.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://doi.org/10.1145/3729397",
"publication_status_note": "ACM FSE published paper; full manuscript not audited",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PUBLISHER_DOI_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1145/3729397",
"previous_chapter_audit_registers": "CH05:PUBLISHER_FULL_HTML_LOCATED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W045PolimiIn the published NDSE2 bibliography
Zhao · 2025
Agents succeed functionally much more often than they satisfy security criteria, making correctness and security separate gates.
Evidence, context & provenance
Evidence boundaryBenchmark selected vulnerability-inducing feature requests.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Benchmarks and construct validity · Vulnerability and secure code
17 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Zhao",
"source": "https://arxiv.org/abs/2512.03262",
"contribution": "Agents succeed functionally much more often than they satisfy security criteria, making correctness and security separate gates.",
"limitation": "Benchmark selected vulnerability-inducing feature requests.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:PDF_PAGE_RELEVANT; CH07:FULLTEXT_HTML",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\", \"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W046PolimiIn the published NDSE2 bibliography
Perry · 2023
On several security tasks the 2022-era assistant lowers the share of secure submissions while increasing confidence.
Evidence, context & provenance
Evidence boundarySmall predominantly student sample; old Codex model.
Source type, as recorded: Controlled human or field experiment
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Developer workflow and adoption · Automation bias and trust · Agent failure and escalation · Vulnerability and secure code · Skills formation and teaching · Dataset provenance and dependency
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Do Users Write More Insecure Code with AI Assistants?",
"kind": "paper",
"group": "Code quality & security",
"year": "2023",
"lead_author": "Perry",
"source": "https://arxiv.org/abs/2211.03622",
"contribution": "On several security tasks the 2022-era assistant lowers the share of secure submissions while increasing confidence.",
"limitation": "Small predominantly student sample; old Codex model.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2023",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:PDF_PAGE_RELEVANT",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W047PolimiIn the published NDSE2 bibliography
Sandoval · 2023
Assisted and control groups exhibit similar security-critical bug rates for one C linked-list implementation.
Evidence, context & provenance
Evidence boundaryOne low-level task; cannot generalize or negate other studies.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Vulnerability and secure code · Skills formation and teaching · Evaluation methodology and baselines
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Lost at C: A User Study on the Security Implications of Large Language Model Code Assistants",
"kind": "paper",
"group": "Code quality & security",
"year": "2023",
"lead_author": "Sandoval",
"source": "https://arxiv.org/abs/2208.09727",
"contribution": "Assisted and control groups exhibit similar security-critical bug rates for one C linked-list implementation.",
"limitation": "One low-level task; cannot generalize or negate other studies.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2023",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W048PolimiIn the published NDSE2 bibliography
Pearce · 2022
Copilot generates vulnerable implementations in deliberately security-sensitive prompts, motivating deterministic security checks.
Evidence, context & provenance
Evidence boundaryCurated weak scenarios; 2021 model and no realistic prevalence estimate.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Benchmarks and construct validity · Vulnerability and secure code
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions",
"kind": "paper",
"group": "Code quality & security",
"year": "2022",
"lead_author": "Pearce",
"source": "https://arxiv.org/abs/2108.09293",
"contribution": "Copilot generates vulnerable implementations in deliberately security-sensitive prompts, motivating deterministic security checks.",
"limitation": "Curated weak scenarios; 2021 model and no realistic prevalence estimate.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2022",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W049PolimiIn the published NDSE2 bibliography
Churilov · 2026
Modern frontier models hallucinate fewer packages, but shared plausible unregistered names remain exploitable.
Evidence, context & provenance
Evidence boundaryPackage frequency is a model-and-task dependent measurement.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Convert the reported risk into a static, behavioral or security acceptance gate.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Dependencies and agent supply chain · Evaluation methodology and baselines · Dataset provenance and dependency
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Range Shrinks, the Threat Remains: Re-evaluating LLM Package Hallucinations on the 2026 Frontier-Model Cohort",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Churilov",
"source": "https://arxiv.org/abs/2605.17062",
"contribution": "Modern frontier models hallucinate fewer packages, but shared plausible unregistered names remain exploitable.",
"limitation": "Package frequency is a model-and-task dependent measurement.",
"manual_use": "Convert the reported risk into a static, behavioral or security acceptance gate.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W050PolimiIn the published NDSE2 bibliography
Deng · 2026
Most audited apps had detectable vulnerabilities; basic prompting only partially prevents replayed mistakes.
Evidence, context & provenance
Evidence boundarySelective sampling; not representative prevalence of all AI-created apps.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Vulnerability and secure code · System understanding and rationale
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Understanding the (In)Security of Vibe-Coded Applications",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Deng",
"source": "https://arxiv.org/abs/2606.23130",
"contribution": "Most audited apps had detectable vulnerabilities; basic prompting only partially prevents replayed mistakes.",
"limitation": "Selective sampling; not representative prevalence of all AI-created apps.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W051PolimiIn the published NDSE2 bibliography
Fang · 2026
No broad degradation on observed quality signals, but developers anticipate worse maintainability of others' AI output.
Evidence, context & provenance
Evidence boundaryPerceived maintainability is not measured effort; sampling differs from negative studies.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Automation bias and trust · Evaluation methodology and baselines
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "From Conversation to Contribution: Characterizing Coding Agent in Open-Source Software",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Fang",
"source": "https://arxiv.org/abs/2607.05677",
"contribution": "No broad degradation on observed quality signals, but developers anticipate worse maintainability of others' AI output.",
"limitation": "Perceived maintainability is not measured effort; sampling differs from negative studies.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W052PolimiIn the published NDSE2 bibliography
Improta · 2026
Generated functions are shorter but security effects vary across Python Java and C; normalize size before judging complexity.
Evidence, context & provenance
Evidence boundaryPreprint; function-level comparisons not entire codebases.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Technical debt and complexity · Vulnerability and secure code · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "What is the Difference Between Me and You? Benchmarking the Quality Gap Between Human-Written and AI-Generated Code",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Improta",
"source": "https://arxiv.org/abs/2609.12708",
"contribution": "Generated functions are shorter but security effects vary across Python Java and C; normalize size before judging complexity.",
"limitation": "Preprint; function-level comparisons not entire codebases.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:TARGETED_FULL_TEXT_PAGES_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W053PolimiIn the published NDSE2 bibliography
Kapner · 2026
Exposed execution grants, versionless MCP servers, and overbroad skill approvals make harness configuration an attack surface.
Evidence, context & provenance
Evidence boundaryStatic configuration exposure does not prove exploitation.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Prompt injection and malicious tools · Dependencies and agent supply chain · Skills formation and teaching
15 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Scanning the Harness: Configuration Exposures in AI Coding-Agent Supply Chains",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Kapner",
"source": "https://arxiv.org/abs/2609.07360",
"contribution": "Exposed execution grants, versionless MCP servers, and overbroad skill approvals make harness configuration an attack surface.",
"limitation": "Static configuration exposure does not prove exploitation.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:PDF_PAGE_RELEVANT",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W054PolimiIn the published NDSE2 bibliography
Kraishan · 2026
Post-merge revert behavior varies across named agents; aggregate agent-vs-human averages hide this heterogeneity.
Evidence, context & provenance
Evidence boundaryReverts inferred from messages and security smells from regex.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Vulnerability and secure code · Evaluation methodology and baselines · Long-running agent execution · Dataset provenance and dependency
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Not All Agents Are Equal: Code Quality and Post-Merge Maintenance Across Five Autonomous Coding Agents in the Wild",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Kraishan",
"source": "https://arxiv.org/abs/2609.17598",
"contribution": "Post-merge revert behavior varies across named agents; aggregate agent-vs-human averages hide this heterogeneity.",
"limitation": "Reverts inferred from messages and security smells from regex.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:TARGETED_FULL_TEXT_PAGES_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C02\", \"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W055PolimiIn the published NDSE2 bibliography
Liu · 2026
Many assistant-authored commits introduce code smells, and some findings survive subsequent maintenance.
Evidence, context & provenance
Evidence boundaryNo comparable human commit baseline; smells not confirmed defects.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Developer workflow and adoption · Technical debt and complexity · Evaluation methodology and baselines
15 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Liu",
"source": "https://arxiv.org/abs/2603.28592",
"contribution": "Many assistant-authored commits introduce code smells, and some findings survive subsequent maintenance.",
"limitation": "No comparable human commit baseline; smells not confirmed defects.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W056PolimiIn the published NDSE2 bibliography
Mao · 2026
AI code has fewer CodeQL alerts and cross-file duplicates in selected units, yet is edited more after introduction.
Evidence, context & provenance
Evidence boundarySelection, granularity and matching differ from repository-level debt studies.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to design independent static-analysis and repair gates; verify residual vulnerability, not just suggested fixes.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Code reuse and duplication · Evaluation methodology and baselines
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "A Large-Scale Comprehensive Measurement of AI-Generated Code in Real-World Repositories",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Mao",
"source": "https://arxiv.org/abs/2603.27130",
"contribution": "AI code has fewer CodeQL alerts and cross-file duplicates in selected units, yet is edited more after introduction.",
"limitation": "Selection, granularity and matching differ from repository-level debt studies.",
"manual_use": "Use to design independent static-analysis and repair gates; verify residual vulnerability, not just suggested fixes.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W057PolimiIn the published NDSE2 bibliography
Mazloomzadeh · 2026
Proposes to compare agent contribution and merge patterns over adoption lifecycle with human PR baselines.
Evidence, context & provenance
Evidence boundaryBibliography explicitly lacks recorded results: no outcome claim warranted.
Source type, as recorded: Missing outcome details
Editorial use. Include in the research inventory but do not cite as support for a measured outcome until the primary results are checked.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Architecture and modularity · Evaluation methodology and baselines · Dataset provenance and dependency
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "How Do AI Coding Agents Contribute to Software Development? An Empirical Study of Agentic Pull Requests",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Mazloomzadeh",
"source": "https://arxiv.org/abs/2607.21832",
"contribution": "Proposes to compare agent contribution and merge patterns over adoption lifecycle with human PR baselines.",
"limitation": "Bibliography explicitly lacks recorded results: no outcome claim warranted.",
"manual_use": "Include in the research inventory but do not cite as support for a measured outcome until the primary results are checked.",
"evidence_class": "Missing outcome details",
"editorial_priority": "VERIFY",
"bibliography_year": "2026",
"original_evidence_class": "Missing outcome details",
"reviewed_evidence_class": "Missing outcome details",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W058PolimiIn the published NDSE2 bibliography
Xu · 2026
Agent adoption correlates with higher code complexity but does not significantly reduce newcomer participation or retention.
Evidence, context & provenance
Evidence boundaryNo actual onboarding time or knowledge measurement.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Developer workflow and adoption · Technical debt and complexity · Onboarding and developer fluency · Evaluation methodology and baselines
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Decoupling Code Complexity from Newcomer Participation: A Causal Study of AI Coding Agent Adoption in OSS",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Xu",
"source": "https://arxiv.org/abs/2607.01810",
"contribution": "Agent adoption correlates with higher code complexity but does not significantly reduce newcomer participation or retention.",
"limitation": "No actual onboarding time or knowledge measurement.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W059PolimiIn the published NDSE2 bibliography
Zhu · 2026
Larger generated outputs are associated with coupling and architectural smells even when tests pass.
Evidence, context & provenance
Evidence boundarySmall smell counts; correlation with LOC not a causal law.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Independent verification and oracles · Benchmarks and construct validity · Technical debt and complexity · Architecture and modularity
15 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI-Generated Smells: An Analysis of Code and Architecture in LLM- and Agent-Driven Development",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Zhu",
"source": "https://arxiv.org/abs/2605.02741",
"contribution": "Larger generated outputs are associated with coupling and architectural smells even when tests pass.",
"limitation": "Small smell counts; correlation with LOC not a causal law.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:NOT_FULL_TEXT_VERIFIED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W060PolimiIn the published NDSE2 bibliography
Liu · 2025
Workspace and retrieved-content injections can induce agents to execute unwanted commands.
Evidence, context & provenance
Evidence boundaryAttack configurations are adversarial, not real-world incidence rates.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Prompt injection and malicious tools
24 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Your AI, My Shell: Demystifying Prompt Injection Attacks on Agentic AI Coding Editors",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Liu",
"source": "https://arxiv.org/abs/2509.22040",
"contribution": "Workspace and retrieved-content injections can induce agents to execute unwanted commands.",
"limitation": "Attack configurations are adversarial, not real-world incidence rates.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:ARXIV_ABSTRACT_VERIFIED; CH06:PRIMARY_ABSTRACT_ONLY",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C06\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W061PolimiIn the published NDSE2 bibliography
Veracode · 2025
Functional model progress does not automatically translate into fewer code security-test failures.
Evidence, context & provenance
Evidence boundaryVendor-created prompts and benchmarks; not peer reviewed.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Benchmarks and construct validity · Agent failure and escalation · Vulnerability and secure code
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "2025 GenAI Code Security Report",
"kind": "paper",
"group": "Code quality & security",
"year": "2025",
"lead_author": "Veracode",
"source": "https://www.veracode.com/wp-content/uploads/2025_GenAI_Code_Security_Report_Final.pdf",
"contribution": "Functional model progress does not automatically translate into fewer code security-test failures.",
"limitation": "Vendor-created prompts and benchmarks; not peer reviewed.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2025",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://www.veracode.com/wp-content/uploads/2025_GenAI_Code_Security_Report_Final.pdf",
"pass7_identity_source_class": "original_report_pdf",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-code-quality-security",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W062PolimiIn the published NDSE2 bibliography
Zhao · 2026
Public vulnerability disclosures with coding-agent provenance show verifiable safety failures in the field.
Evidence, context & provenance
Evidence boundarySource marker coverage is incomplete; denominator not interpretable as risk rate.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Agent failure and escalation · Vulnerability and secure code · Specifications and governance
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Vibe Security Radar",
"kind": "paper",
"group": "Code quality & security",
"year": "2026",
"lead_author": "Zhao",
"source": "https://github.com/HQ1995/vibe-security-radar",
"contribution": "Public vulnerability disclosures with coding-agent provenance show verifiable safety failures in the field.",
"limitation": "Source marker coverage is incomplete; denominator not interpretable as risk rate.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://github.com/HQ1995/vibe-security-radar",
"pass7_identity_source_class": "author_software_repository",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-code-quality-security",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W063PolimiIn the published NDSE2 bibliography
Baltes · 2026
Open-source contributors describe reviewer overload and quality externalities when code generation becomes cheaper than review.
Evidence, context & provenance
Evidence boundaryPublic forum perceptions are not verified costs or defect counts.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles · Economic costs and return
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "An Endless Stream of AI Slop: How Developers Discuss the Burden of AI-Assisted Software Development",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Baltes",
"source": "https://arxiv.org/abs/2603.27249",
"contribution": "Open-source contributors describe reviewer overload and quality externalities when code generation becomes cheaper than review.",
"limitation": "Public forum perceptions are not verified costs or defect counts.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W064PolimiIn the published NDSE2 bibliography
Crupi · 2026
AI comments cover only a small share of human findings and generate much additional low-value review text.
Evidence, context & provenance
Evidence boundaryOne model generation and selected PRs.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Studying Quality Improvements Recommended via Manual and Automated Code Review",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Crupi",
"source": "https://arxiv.org/abs/2602.11925",
"contribution": "AI comments cover only a small share of human findings and generate much additional low-value review text.",
"limitation": "One model generation and selected PRs.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W065PolimiIn the published NDSE2 bibliography
Duma · 2026
Most agent PRs have no recorded human review, but same-repository human PRs also often lack review.
Evidence, context & provenance
Evidence boundaryRecorded GitHub review is not total review rigor; shared AIDev data.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Dataset provenance and dependency
19 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "These Aren't the Reviews You're Looking For: How Humans Review AI-Generated Pull Requests",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Duma",
"source": "https://arxiv.org/abs/2605.02273",
"contribution": "Most agent PRs have no recorded human review, but same-repository human PRs also often lack review.",
"limitation": "Recorded GitHub review is not total review rigor; shared AIDev data.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W066PolimiIn the published NDSE2 bibliography
Gao · 2026
Explanations alter review trust and perceived usefulness; trust must be distinguished from objective agreement.
Evidence, context & provenance
Evidence boundarySmall participant pool and proxy trust metrics.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Automation bias and trust · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Evaluating the Impact of Explainable AI on Trust in AI-Assisted Code Review",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Gao",
"source": "https://arxiv.org/abs/2607.24601",
"contribution": "Explanations alter review trust and perceived usefulness; trust must be distinguished from objective agreement.",
"limitation": "Small participant pool and proxy trust metrics.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://arxiv.org/abs/2607.24601",
"pass7_identity_source_class": "author_preprint",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W067PolimiIn the published NDSE2 bibliography
Khalid · 2026
Users select unsafe suggestions nearly as often as safe ones; editing correlates with fewer vulnerabilities.
Evidence, context & provenance
Evidence boundaryFour selected C tasks; 2027 conference claim relies on current preprint.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Automation bias and trust · Independent verification and oracles · Vulnerability and secure code
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "(Don't) Trust, but (Don't) Verify: Developers' Attention to Security in AI-Generated Code",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Khalid",
"source": "https://arxiv.org/abs/2609.21020",
"contribution": "Users select unsafe suggestions nearly as often as safe ones; editing correlates with fewer vulnerabilities.",
"limitation": "Four selected C tasks; 2027 conference claim relies on current preprint.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W068PolimiIn the published NDSE2 bibliography
Khelifi · 2026
Human interventions occur on fewer agent PRs, but those interventions are often more extensive and guidance-oriented.
Evidence, context & provenance
Evidence boundaryMetadata events cannot certify actual review quality; shared AIDev data.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Agent failure and escalation · Human-agent collaboration · Dataset provenance and dependency
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Behind Agentic Pull Requests: An Empirical Study on Developer Interventions in AI Agent-Authored Pull Requests",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Khelifi",
"source": "https://espace2.etsmtl.ca/id/eprint/34495/",
"contribution": "Human interventions occur on fewer agent PRs, but those interventions are often more extensive and guidance-oriented.",
"limitation": "Metadata events cannot certify actual review quality; shared AIDev data.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://espace2.etsmtl.ca/id/eprint/34495/",
"publication_status_note": "AIDev observational cohort reused; incomplete independence",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_INSTITUTION_PUBLICATION",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1145/3793302.3793586",
"data_snapshot_warning": "Uses AIDev PR records.",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass3_audit_primary_url": "https://espace2.etsmtl.ca/id/eprint/34495/",
"pass3_audit_version": "MSR 2026 published author copy pp.843-844",
"pass3_audit_depth": "ORIGINAL_PAPER_METHODS_AND_TABLE_SELECTED",
"pass3_audit_methods_locators": "Table 1 four human cells versus stated 5869; results tables",
"pass3_audit_sample": "Repository-matched AIDev subset closed PRs; 5869 agent, stated 5869 human",
"pass3_audit_observation": "Human count cells sum to 5853; mismatch 16; interventions 52.17/83.59% reported",
"pass3_audit_boundary": "Unresolved missing records and observational selection; elapsed time is not reviewer labor",
"pass3_audit_original_data_replication": "NO",
"pass3_audit_provenance": "PASS3_PRINTED_EDITION_RECONSTRUCTED_2026-10-10",
"pass3_audit_date": "2026-10-10",
"pass6_date": "2026-10-10",
"pass6_locator": "https://espace2.etsmtl.ca/id/eprint/34495/1/Ouni-A-2026-34495.pdf",
"pass6_inspection_depth": "PUBLISHED_ACM_PROCEEDINGS_TABLE1_P2",
"pass6_page_scope": "p2 Table1 and p1 methods",
"pass6_finding": "HPR 5853 printed vs 5869 described; arithmetic-inferred final cell 261; 52.17/83.59% intervention margins",
"pass6_limitation": "RAW_OBSERVATIONS_NOT_AVAILABLE",
"pass6_original_data_replicated": "NO",
"ndse2_actual_chapters_json": "[\"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W069PolimiIn the published NDSE2 bibliography
Khojah · 2026
The AI label increases fixation time without evidence that defect detection improves.
Evidence, context & provenance
Evidence boundaryIdentical human-origin code; injected-defect detection not assessed.
Source type, as recorded: Controlled human or field experiment
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Automation bias and trust · Independent verification and oracles · Human cognitive load
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Same Scrutiny, More Time: Eye Tracking Insights into Reviewing LLM-Labelled Code",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Khojah",
"source": "https://arxiv.org/abs/2606.26505",
"contribution": "The AI label increases fixation time without evidence that defect detection improves.",
"limitation": "Identical human-origin code; injected-defect detection not assessed.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:PRIMARY_ABSTRACT_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W070PolimiIn the published NDSE2 bibliography
Olewicki · 2026
AI review comments have modest acceptance and must be filtered for relevance, especially functional advice.
Evidence, context & provenance
Evidence boundaryAcceptance is a process proxy, not defect prevention.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · AI as reviewer
3 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Impact of an LLM-based Review Assistant in Practice: A Mixed Open-/Closed-source Case Study",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Olewicki",
"source": "https://arxiv.org/abs/2411.07091",
"contribution": "AI review comments have modest acceptance and must be filtered for relevance, especially functional advice.",
"limitation": "Acceptance is a process proxy, not defect prevention.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W071PolimiIn the published NDSE2 bibliography
Peralta · 2026
Rejection categories include agent failure, project workflow mismatch and unspecified reasons; distinguish causes before retrying.
Evidence, context & provenance
Evidence boundaryExplanations may be inferred from inconsistent PR records; shared AIDev data.
Source type, as recorded: AIDEV_DERIVED_OBSERVATIONAL_STRATIFIED_CODING
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Agent failure and escalation · Specifications and governance · Dataset provenance and dependency
24 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Why Are Agentic Pull Requests Merged or Rejected? An Empirical Study",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Peralta",
"source": "https://arxiv.org/html/2605.22534",
"contribution": "Rejection categories include agent failure, project workflow mismatch and unspecified reasons; distinguish causes before retrying.",
"limitation": "Explanations may be inferred from inconsistent PR records; shared AIDev data.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "AIDEV_DERIVED_OBSERVATIONAL_STRATIFIED_CODING",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "AIDEV_DERIVED_OBSERVATIONAL_STRATIFIED_CODING",
"full_manuscript_audited": "yes_selected_methods",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/html/2605.22534",
"publication_status_note": "2026 MSR paper, Sections 3-6; explicitly builds on AIDev",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "FULLTEXT_HTML_METHODS_TABLES",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_MANUSCRIPT_SECTIONS",
"audit_locator": "MSR 2026 arXiv 2605.22534 html Sections 3-6 and Tables 1-3",
"doi": "10.1145/3793302.3793575",
"data_snapshot_warning": "Uses AIDev PR records (MSR 2026 Sect. 3), 9799 human-reviewed filtered PRs, 717 stratified manually assessed.",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C04\", \"C:C07\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W072PolimiIn the published NDSE2 bibliography
Selvanayagam · 2026
AI agents are already reviewing other agents, but automation density says nothing about independent error detection.
Evidence, context & provenance
Evidence boundaryDistinct AI-to-AI actors rare, and quality not measured.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI-to-AI Code Reviews of GitHub Pull Requests",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Selvanayagam",
"source": "https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.ESEM.2026.75",
"contribution": "AI agents are already reviewing other agents, but automation density says nothing about independent error detection.",
"limitation": "Distinct AI-to-AI actors rare, and quality not measured.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.ESEM.2026.75",
"pass7_identity_source_class": "peer_reviewed_open_access",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W073PolimiIn the published NDSE2 bibliography
Singh · 2026
Complexity makes reviewers more likely to accept wrong changes; AI labeling alone has no significant outcome effect.
Evidence, context & provenance
Evidence boundaryControlled decisions rather than production incidents.
Source type, as recorded: Controlled human or field experiment
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Review capacity and maintainers · Technical debt and complexity · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Effect of Complexity and Provenance on Code Review Decisions: Evidence from a Controlled Experiment",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Singh",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Complexity makes reviewers more likely to accept wrong changes; AI labeling alone has no significant outcome effect.",
"limitation": "Controlled decisions rather than production incidents.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W074PolimiIn the published NDSE2 bibliography
Tantithamthavorn · 2026
AI-assisted review associates with shorter PR cycles and fewer human comments.
Evidence, context & provenance
Evidence boundaryNo defect-rate or deployed-incident endpoint.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "RovoDev Code Reviewer: A Large-Scale Online Evaluation of LLM-based Code Review Automation at Atlassian",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Tantithamthavorn",
"source": "https://arxiv.org/abs/2601.01129",
"contribution": "AI-assisted review associates with shorter PR cycles and fewer human comments.",
"limitation": "No defect-rate or deployed-incident endpoint.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:PRIMARY_ABSTRACT_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W075PolimiIn the published NDSE2 bibliography
Tsui · 2026
Models often correct an externally attributed mistake more successfully than their own.
Evidence, context & provenance
Evidence boundaryNon-programming tasks; transfer to SWE is a hypothesis.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Benchmarks and construct validity · Agent failure and escalation
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Self-Correction Bench: Uncovering and Addressing the Self-Correction Blind Spot in LLMs",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Tsui",
"source": "https://arxiv.org/abs/2507.02778",
"contribution": "Models often correct an externally attributed mistake more successfully than their own.",
"limitation": "Non-programming tasks; transfer to SWE is a hypothesis.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH07:FULLTEXT_HTML",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W076PolimiIn the published NDSE2 bibliography
Zeng · 2026
Best model review approaches show limited precision and recall for verified review findings.
Evidence, context & provenance
Evidence boundaryBenchmark is model-only, and human collaborative settings differ.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles · Benchmarks and construct validity · Employment and early-career roles
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "SWR-Bench: Assessing LLM Performance in Real-World Code Review Comment Generation",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Zeng",
"source": "https://arxiv.org/abs/2509.01494",
"contribution": "Best model review approaches show limited precision and recall for verified review findings.",
"limitation": "Benchmark is model-only, and human collaborative settings differ.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:PRIMARY_ABSTRACT_CHECKED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W077PolimiIn the published NDSE2 bibliography
Zhao · 2026
Tests generated directly from defective implementations can encode the bug; an independent specification reduces error propagation.
Evidence, context & provenance
Evidence boundarySpecific Java defect benchmark and prompt setup.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Benchmarks and construct validity · Specifications and governance · Evaluation methodology and baselines · Human-agent collaboration
32 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Evaluating and Mitigating the Misguidance Effect of Buggy Code in LLM-Generated Unit Tests",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Zhao",
"source": "https://arxiv.org/abs/2607.22883",
"contribution": "Tests generated directly from defective implementations can encode the bug; an independent specification reduces error propagation.",
"limitation": "Specific Java defect benchmark and prompt setup.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "2026",
"publication_source_url": "https://doi.org/10.1145/3832204",
"publication_status_note": "Proceedings of the ACM on Software Engineering / ISSTA 2026",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH04:ABSTRACT_AND_REPOSITORY; CH07:PRIMARY_PDF_PAGE_CHECKED",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C04\", \"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W078PolimiIn the published NDSE2 bibliography
Chen · 2025
Model and training-data similarity influence judgments, complicating same-model evaluator independence.
Evidence, context & provenance
Evidence boundaryText-based judgment tasks, not code reviews.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · AI as reviewer · Skills formation and teaching · Evaluation methodology and baselines
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Beyond the Surface: Measuring Self-Preference in LLM Judgments",
"kind": "paper",
"group": "Verification & review",
"year": "2025",
"lead_author": "Chen",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Model and training-data similarity influence judgments, complicating same-model evaluator independence.",
"limitation": "Text-based judgment tasks, not code reviews.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH07:FULLTEXT_HTML_PUBLISHER",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W079PolimiIn the published NDSE2 bibliography
Kim · 2025
Errors can be shared across models, so a second model is not automatically an independent checker.
Evidence, context & provenance
Evidence boundaryGeneral LLM datasets; may not predict specific coding errors.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Benchmarks and construct validity · Dataset provenance and dependency
20 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Correlated Errors in Large Language Models",
"kind": "paper",
"group": "Verification & review",
"year": "2025",
"lead_author": "Kim",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Errors can be shared across models, so a second model is not automatically an independent checker.",
"limitation": "General LLM datasets; may not predict specific coding errors.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "Generic model error correlation (transfer hypothesis)",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH07:FULLTEXT_HTML_PUBLISHER; CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "3",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass3_audit_primary_url": "https://arxiv.org/html/2506.07962",
"pass3_audit_version": "arXiv research preprint 2025",
"pass3_audit_depth": "ORIGINAL_HTML_SELECTED_DESIGN",
"pass3_audit_methods_locators": "Data/Methods and conditional-error calculation",
"pass3_audit_sample": "349 models and 12032 MCQs on one leaderboard; 71 models/14042 questions on HELM",
"pass3_audit_observation": "~60% same wrong answer conditional on both models wrong on HELM",
"pass3_audit_boundary": "Not code-review or defect co-detection; cannot set checker correlation rate",
"pass3_audit_original_data_replication": "NO",
"pass3_audit_provenance": "PASS3_PRINTED_EDITION_RECONSTRUCTED_2026-10-10",
"pass3_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C07\", \"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W080PolimiIn the published NDSE2 bibliography
Ramler · 2025
LLM-supported unit-test writing detects more bugs than an earlier unassisted cohort.
Evidence, context & provenance
Evidence boundaryNon-random historical controls and learner population.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Economic costs and return · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Unit Testing Past vs. Present: Examining LLMs' Impact on Defect Detection and Efficiency",
"kind": "paper",
"group": "Verification & review",
"year": "2025",
"lead_author": "Ramler",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "LLM-supported unit-test writing detects more bugs than an earlier unassisted cohort.",
"limitation": "Non-random historical controls and learner population.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH07:FULLTEXT_HTML_PUBLISHER",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W081PolimiIn the published NDSE2 bibliography
Tufano · 2025
Starting with LLM comments consumes more time without finding more major seeded defects.
Evidence, context & provenance
Evidence boundarySmall controlled sample; one system, but true defects observable.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Skills formation and teaching · Evaluation methodology and baselines · Dataset provenance and dependency
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Deep Learning-based Code Reviews: A Paradigm Shift or a Double-Edged Sword?",
"kind": "paper",
"group": "Verification & review",
"year": "2025",
"lead_author": "Tufano",
"source": "https://arxiv.org/abs/2411.11401",
"contribution": "Starting with LLM comments consumes more time without finding more major seeded defects.",
"limitation": "Small controlled sample; one system, but true defects observable.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W082PolimiIn the published NDSE2 bibliography
Panickssery · 2024
LLM judges recognize and favor their own prose generations, cautioning against self-grading loops.
Evidence, context & provenance
Evidence boundarySummarization benchmark, not code.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. AI as reviewer · Benchmarks and construct validity
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "LLM Evaluators Recognize and Favor Their Own Generations",
"kind": "paper",
"group": "Verification & review",
"year": "2024",
"lead_author": "Panickssery",
"source": "https://arxiv.org/abs/2404.13076",
"contribution": "LLM judges recognize and favor their own prose generations, cautioning against self-grading loops.",
"limitation": "Summarization benchmark, not code.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2024",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "Generic LLM-as-a-judge preference (transfer hypothesis)",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH07:FULLTEXT_HTML",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W083PolimiIn the published NDSE2 bibliography
Tang · 2024
Disclosing machine authorship increases inspection and improves code validation and repair in a small experiment.
Evidence, context & provenance
Evidence boundaryTwenty-eight participants; different from passive label studies.
Source type, as recorded: Controlled human or field experiment
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Specifications and governance · Human cognitive load · Ownership and knowledge transfer
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "A Study on Developer Behaviors for Validating and Repairing LLM-Generated Code Using Eye Tracking and IDE Actions",
"kind": "paper",
"group": "Verification & review",
"year": "2024",
"lead_author": "Tang",
"source": "https://arxiv.org/abs/2405.16081",
"contribution": "Disclosing machine authorship increases inspection and improves code validation and repair in a small experiment.",
"limitation": "Twenty-eight participants; different from passive label studies.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2024",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W084PolimiIn the published NDSE2 bibliography
Wataoka · 2024
Judges favor model-familiar outputs; stylistic similarity can masquerade as quality.
Evidence, context & provenance
Evidence boundaryDialogue judgments, not production software.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · AI as reviewer
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Self-Preference Bias in LLM-as-a-Judge",
"kind": "paper",
"group": "Verification & review",
"year": "2024",
"lead_author": "Wataoka",
"source": "https://arxiv.org/abs/2410.21819",
"contribution": "Judges favor model-familiar outputs; stylistic similarity can masquerade as quality.",
"limitation": "Dialogue judgments, not production software.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2024",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W085PolimiIn the published NDSE2 bibliography
Dratsch · 2023
Even expert readers follow wrong AI suggestions, supporting a general need for calibrated skepticism.
Evidence, context & provenance
Evidence boundaryClinical imaging analogy, not direct SWE evidence.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Automation bias and trust
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance",
"kind": "paper",
"group": "Verification & review",
"year": "2023",
"lead_author": "Dratsch",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Even expert readers follow wrong AI suggestions, supporting a general need for calibrated skepticism.",
"limitation": "Clinical imaging analogy, not direct SWE evidence.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2023",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "Clinical mammography automation bias, not source review",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass6_primary_locator": "https://pubs.rsna.org/doi/10.1148/radiol.222176",
"pass6_metadata_depth": "PUBLISHED_PEER_REVIEWED_JOURNAL_DOI",
"pass6_supporting_locator": "https://pubmed.ncbi.nlm.nih.gov/37129490/",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W086PolimiIn the published NDSE2 bibliography
Hijazi · 2023
Physiological signals may be correlated with quality of code review rather than just speed.
Evidence, context & provenance
Evidence boundarySample size in bibliography unverified; invasive and nonstandard.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles · Human cognitive load · System understanding and rationale · Dataset provenance and dependency
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Quality Evaluation of Modern Code Reviews Through Intelligent Biometric Program Comprehension",
"kind": "paper",
"group": "Verification & review",
"year": "2023",
"lead_author": "Hijazi",
"source": "https://doi.org/10.1109/TSE.2022.3158543",
"contribution": "Physiological signals may be correlated with quality of code review rather than just speed.",
"limitation": "Sample size in bibliography unverified; invasive and nonstandard.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2023",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1109/TSE.2022.3158543",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W087PolimiIn the published NDSE2 bibliography
Gaube · 2021
Incorrect AI-attributed advice induces substantial errors even when professionals are experienced.
Evidence, context & provenance
Evidence boundaryClinical advice, not code review.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Do as AI say: susceptibility in deployment of clinical decision-aids",
"kind": "paper",
"group": "Verification & review",
"year": "2021",
"lead_author": "Gaube",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Incorrect AI-attributed advice induces substantial errors even when professionals are experienced.",
"limitation": "Clinical advice, not code review.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2021",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "Clinical decision-aid behavior, not source review",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass6_primary_locator": "https://www.nature.com/articles/s41746-021-00385-9",
"pass6_metadata_depth": "PUBLISHED_PEER_REVIEWED_JOURNAL_DOI",
"pass6_supporting_locator": "https://pmc.ncbi.nlm.nih.gov/articles/PMC7896064/",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W088PolimiIn the published NDSE2 bibliography
Bertram · 2020
Source labeling changes gaze and subjective ratings without demonstrating accuracy changes.
Evidence, context & provenance
Evidence boundaryTen students; predates current code LLMs.
Source type, as recorded: Controlled human or field experiment
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Review capacity and maintainers · Automation bias and trust · Human cognitive load · Skills formation and teaching
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Trustworthiness Perceptions in Code Review: An Eye-tracking Study",
"kind": "paper",
"group": "Verification & review",
"year": "2020",
"lead_author": "Bertram",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Source labeling changes gaze and subjective ratings without demonstrating accuracy changes.",
"limitation": "Ten students; predates current code LLMs.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2020",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W089PolimiIn the published NDSE2 bibliography
Nourani · 2020
Expertise improves error recognition and trust calibration; initial failures shape later trust.
Evidence, context & provenance
Evidence boundaryArthropod classification, not code.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Automation bias and trust · Agent failure and escalation
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Role of Domain Expertise in User Trust and the Impact of First Impressions with Intelligent Systems",
"kind": "paper",
"group": "Verification & review",
"year": "2020",
"lead_author": "Nourani",
"source": "https://ojs.aaai.org/index.php/HCOMP/article/view/7469",
"contribution": "Expertise improves error recognition and trust calibration; initial failures shape later trust.",
"limitation": "Arthropod classification, not code.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2020",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "Non-code classifier trust experiment",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://ojs.aaai.org/index.php/HCOMP/article/view/7469",
"pass7_identity_source_class": "publisher_record",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W090PolimiIn the published NDSE2 bibliography
Floyd · 2017
Code and prose review rely on distinguishable cognitive processes influenced by expertise.
Evidence, context & provenance
Evidence boundaryNeural classification does not measure review effectiveness in production.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Human cognitive load · System understanding and rationale · Economic costs and return · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Decoding the Representation of Code in the Brain: An fMRI Study of Code Review and Expertise",
"kind": "paper",
"group": "Verification & review",
"year": "2017",
"lead_author": "Floyd",
"source": "https://doi.org/10.1109/ICSE.2017.24",
"contribution": "Code and prose review rely on distinguishable cognitive processes influenced by expertise.",
"limitation": "Neural classification does not measure review effectiveness in production.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2017",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1109/ICSE.2017.24",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W091PolimiIn the published NDSE2 bibliography
Lyell · 2017
Bias can arise in single verification tasks and increases with difficult-to-check suggestions.
Evidence, context & provenance
Evidence boundaryLiterature not directly about code.
Source type, as recorded: Literature synthesis
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Automation bias and trust · Independent verification and oracles · Technical debt and complexity
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Automation bias and verification complexity: a systematic review",
"kind": "paper",
"group": "Verification & review",
"year": "2017",
"lead_author": "Lyell",
"source": "https://pmc.ncbi.nlm.nih.gov/articles/PMC7651899/",
"contribution": "Bias can arise in single verification tasks and increases with difficult-to-check suggestions.",
"limitation": "Literature not directly about code.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Literature synthesis",
"editorial_priority": "B",
"bibliography_year": "2017",
"original_evidence_class": "Literature synthesis",
"reviewed_evidence_class": "Literature synthesis",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "Cross-domain automation bias systematic review",
"publication_year_from_independent_source": "",
"publication_source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC7651899/",
"publication_status_note": "Systematic review of decision-support bias in healthcare/human factors; no direct AI software experiment",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "INDEXED_PRIMARY_JOURNAL_TEXT",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1093/jamia/ocw105",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W092PolimiIn the published NDSE2 bibliography
Parasuraman · 2010
Automation-induced attention changes persist even among experienced professionals, motivating active verification design.
Evidence, context & provenance
Evidence boundaryLegacy automation settings; extrapolation to coding agents must be tested.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · System understanding and rationale
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Complacency and Bias in Human Use of Automation: An Attentional Integration",
"kind": "paper",
"group": "Verification & review",
"year": "2010",
"lead_author": "Parasuraman",
"source": "https://doi.org/10.1177/0018720810376055",
"contribution": "Automation-induced attention changes persist even among experienced professionals, motivating active verification design.",
"limitation": "Legacy automation settings; extrapolation to coding agents must be tested.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2010",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1177/0018720810376055",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W093PolimiIn the published NDSE2 bibliography
Adams · 2026
Risk-ranked PR changes may permit selective automation and reserve reviewer time for consequential changes.
Evidence, context & provenance
Evidence boundaryProduction risk calibration requires independent defect outcomes and monitoring.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles · Vulnerability and secure code · Economic costs and return
15 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Automating Low-Risk Code Review at Meta: RADAR, Risk Calibration, and Review Efficiency",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Adams",
"source": "https://arxiv.org/abs/2605.30208",
"contribution": "Risk-ranked PR changes may permit selective automation and reserve reviewer time for consequential changes.",
"limitation": "Production risk calibration requires independent defect outcomes and monitoring.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/abs/2605.30208",
"publication_status_note": "Original preprint; no full evaluation replicated",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PREPRINT_PRIMARY_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W094PolimiIn the published NDSE2 bibliography
Afroz · 2026
Influx of machine-generated submissions can increase maintainer load and lower merge propensity.
Evidence, context & provenance
Evidence boundaryAttribution and matching complexity; context-dependent.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Agent failure and escalation · Technical debt and complexity · System understanding and rationale · Agent context and persistent memory
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI Slop is DDoSing Open Source: Understanding the Impact of AI-Generated Contributions on Open Source Sustainability",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Afroz",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Influx of machine-generated submissions can increase maintainer load and lower merge propensity.",
"limitation": "Attribution and matching complexity; context-dependent.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING; CH11:ORIGINAL_PDF_METHODS_AND_TABLES",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W095PolimiIn the published NDSE2 bibliography
Bhati · 2026
Proposes analyzing the whole agentic lifecycle across verification reliability and cost.
Evidence, context & provenance
Evidence boundarySynthesis only, no new outcome dataset.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Architecture and modularity · Skills formation and teaching · Economic costs and return · Dataset provenance and dependency
34 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Beyond Code Generation: Reliability, Verification, and Cost Economics in the Agentic Software Development Lifecycle",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Bhati",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Proposes analyzing the whole agentic lifecycle across verification reliability and cost.",
"limitation": "Synthesis only, no new outcome dataset.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH07:FULLTEXT_HTML; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "3",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass3_audit_primary_url": "https://arxiv.org/html/2609.04681",
"pass3_audit_version": "arXiv research synthesis September 2026",
"pass3_audit_depth": "PRIMARY_SYNTHESIS_CLASSIFICATION",
"pass3_audit_methods_locators": "Abstract and evidence citations",
"pass3_audit_sample": "No independent empirical sample",
"pass3_audit_observation": "Discusses verification/cost; cites W003 prior field experiment",
"pass3_audit_boundary": "Synthesis cannot become an independent replication of W003",
"pass3_audit_original_data_replication": "NO",
"pass3_audit_provenance": "PASS3_PRINTED_EDITION_RECONSTRUCTED_2026-10-10",
"pass3_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C07\", \"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W096PolimiIn the published NDSE2 bibliography
Cynthia · 2026
Developer resolution rates differ across reviewing agents and ignore comments that conflict with intended design.
Evidence, context & provenance
Evidence boundaryResolution is not necessarily acceptance or correctness.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Review capacity and maintainers · AI as reviewer · Independent verification and oracles · System understanding and rationale
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Go Home Copilot, You're Drunk: Understanding Developer Responses to Agent-Generated Code Review Comments",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Cynthia",
"source": "https://arxiv.org/abs/2607.21997",
"contribution": "Developer resolution rates differ across reviewing agents and ignore comments that conflict with intended design.",
"limitation": "Resolution is not necessarily acceptance or correctness.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W097PolimiIn the published NDSE2 bibliography
Kaufman · 2026
Programmers inadequately spot bad AI assertions while remaining confident, questioning model-generated test oracles.
Evidence, context & provenance
Evidence boundaryShort Python assertions and 86 participants.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. AI as reviewer · Automation bias and trust · Independent verification and oracles · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Programmers Are Poor and Overconfident Judges of LLM-Generated Assertions",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Kaufman",
"source": "https://arxiv.org/abs/2607.08885",
"contribution": "Programmers inadequately spot bad AI assertions while remaining confident, questioning model-generated test oracles.",
"limitation": "Short Python assertions and 86 participants.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH04:ABSTRACT; CH07:PRIMARY_PDF_ABSTRACT_PAGE_CHECKED",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C04\", \"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W098PolimiIn the published NDSE2 bibliography
Lin · 2026
Developers reject many AI reviewer comments, chiefly for false positives and irrelevance.
Evidence, context & provenance
Evidence boundaryOnly comments that received replies, inducing response-selection bias.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · AI as reviewer
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Is Agentic Code Review Helpful? Mining Developers' Feedback to CodeRabbit Reviews in the Wild",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Lin",
"source": "https://arxiv.org/abs/2607.03316",
"contribution": "Developers reject many AI reviewer comments, chiefly for false positives and irrelevance.",
"limitation": "Only comments that received replies, inducing response-selection bias.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W099PolimiIn the published NDSE2 bibliography
Maddila · 2026
Comparing diff content to stated intent can help authors find unintended deviations.
Evidence, context & provenance
Evidence boundarySmall reported effect; defect denominator absent.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · AI as reviewer · Agent failure and escalation · Specifications and governance · Long-running agent execution
28 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "From Code Review to Code Critique: Intent, Drift, and Spotlight for AI-Generated Diffs at Scale",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Maddila",
"source": "https://arxiv.org/abs/2607.29516",
"contribution": "Comparing diff content to stated intent can help authors find unintended deviations.",
"limitation": "Small reported effect; defect denominator absent.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH04:ABSTRACT; CH07:FULLTEXT_HTML; CH08:FULL_TEXT_AUDIT_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "4",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass3_audit_primary_url": "https://arxiv.org/html/2607.29516v1",
"pass3_audit_version": "arXiv:2607.29516v1 (2026-08-14)",
"pass3_audit_depth": "ORIGINAL_HTML_SELECTED_METHODS_AND_TABLES",
"pass3_audit_methods_locators": "Intent Table 3; drift Table 4; rollout Table 8",
"pass3_audit_sample": "121 intent cases, 118 balanced drift cases; opt-in enterprise rollout",
"pass3_audit_observation": "Intent F1 .860/.844; 52191/8089 tokens; drift exact 64.4%; reversions .33/.17%",
"pass3_audit_boundary": "Selected self-attesting cohort not randomized; differing outcomes and sample sizes",
"pass3_audit_original_data_replication": "NO",
"pass3_audit_provenance": "PASS3_PRINTED_EDITION_RECONSTRUCTED_2026-10-10",
"pass3_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C04\", \"C:C07\", \"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W100PolimiIn the published NDSE2 bibliography
Monperrus · 2026
Challenges human-only review and motivates automatic verification research.
Evidence, context & provenance
Evidence boundaryNo empirical testing; should be presented as contested prediction.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles
14 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The End of Code Review: Coding Agents Supersede Human Inspection",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Monperrus",
"source": "https://arxiv.org/abs/2606.13175",
"contribution": "Challenges human-only review and motivates automatic verification research.",
"limitation": "No empirical testing; should be presented as contested prediction.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W101PolimiIn the published NDSE2 bibliography
Purnachandra Reddy · 2026
Models can explain a semantic migration rule and still miss behavior changes in their own port.
Evidence, context & provenance
Evidence boundarySmall Python migration traps; individual model variation high.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. AI as reviewer · Independent verification and oracles · Benchmarks and construct validity · Agent failure and escalation · Agent context and persistent memory
17 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Articulate but Wrong: Self-Review Failures in LLM-Based Code Modernization",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Purnachandra Reddy",
"source": "https://arxiv.org/abs/2605.21537",
"contribution": "Models can explain a semantic migration rule and still miss behavior changes in their own port.",
"limitation": "Small Python migration traps; individual model variation high.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH07:FULLTEXT_HTML",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W102PolimiIn the published NDSE2 bibliography
Xiang · 2026
Independent model review helps some model orderings but can worsen an already strong implementation.
Evidence, context & provenance
Evidence boundaryTwo models, narrow benchmark, multiple comparisons.
Source type, as recorded: Controlled benchmark; complete-case subset, not human field trial
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles · Benchmarks and construct validity · Evaluation methodology and baselines
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Cross-Model LLM Code Review: Should You Use Claude to Review Codex or Vice Versa?",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Xiang",
"source": "https://arxiv.org/abs/2607.21656",
"contribution": "Independent model review helps some model orderings but can worsen an already strong implementation.",
"limitation": "Two models, narrow benchmark, multiple comparisons.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Controlled benchmark; complete-case subset, not human field trial",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH07:PRIMARY_PDF_PAGE_CHECKED; CH08:PRIMARY_ABSTRACT_CHECKED",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass2_source_locator": "https://arxiv.org/html/2607.21656",
"pass2_source_version": "arXiv 2607.21656v1 2026-07-22",
"pass2_inspection_scope": "Sections 3.1--3.4 4.1--4.4 Tables 2 and 4",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NO",
"pass2_research_design": "Controlled benchmark with complete-case task selection",
"pass2_critical_limitation": "No human review; model cannot run tests during review; complete cases and specific model pricing; not repo representative",
"pass2_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W103PolimiIn the published NDSE2 bibliography
Yabesi · 2026
Deadline and perceived seniority alter visual review effort more than final decisions.
Evidence, context & provenance
Evidence boundarySource labels manipulated; performance unchanged.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Automation bias and trust · Human cognitive load · Employment and early-career roles · Evaluation methodology and baselines
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "An Eye for Trust: An Exploration of Developers' Trust Perceptions Through Urgency and Reputation",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Yabesi",
"source": "https://arxiv.org/abs/2604.08713",
"contribution": "Deadline and perceived seniority alter visual review effort more than final decisions.",
"limitation": "Source labels manipulated; performance unchanged.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W104PolimiIn the published NDSE2 bibliography
Ye · 2026
Most users miss covert exfiltration behavior, and alerts alone often fail to prevent acceptance.
Evidence, context & provenance
Evidence boundarySimulated malicious-agent setting, not baseline prevalence.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Agent failure and escalation · Prompt injection and malicious tools · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Coding with Enemy: Can Human Developers Detect AI Agent Sabotage?",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Ye",
"source": "https://arxiv.org/abs/2606.05647",
"contribution": "Most users miss covert exfiltration behavior, and alerts alone often fail to prevent acceptance.",
"limitation": "Simulated malicious-agent setting, not baseline prevalence.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH06:SOURCE_LOCATOR_ONLY; CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C06\", \"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W105PolimiIn the published NDSE2 bibliography
Yu · 2026
Reviewers approve more agent PRs over time, but changes in detection of defects remain unknown.
Evidence, context & provenance
Evidence boundarySelection, learning and project dynamics can mimic habituation.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Automation bias and trust · Skills formation and teaching · Dataset provenance and dependency
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Beyond Lexical Metrics: Sentence-Embedding Detection of Reviewer Habituation in AI Code Review",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Yu",
"source": "https://arxiv.org/abs/2609.06213",
"contribution": "Reviewers approve more agent PRs over time, but changes in detection of defects remain unknown.",
"limitation": "Selection, learning and project dynamics can mimic habituation.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W106PolimiIn the published NDSE2 bibliography
Zhang · 2026
Multiple agent reviewers fail most human-identified actionable review checks, especially architectural issues.
Evidence, context & provenance
Evidence boundaryHuman ground truth built into benchmark by construction.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Benchmarks and construct validity · Architecture and modularity
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Code Review Agent Benchmark (c-CRAB)",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Zhang",
"source": "https://arxiv.org/abs/2603.23448",
"contribution": "Multiple agent reviewers fail most human-identified actionable review checks, especially architectural issues.",
"limitation": "Human ground truth built into benchmark by construction.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W107PolimiIn the published NDSE2 bibliography
Zhong · 2026
Human-on-agent discussions are longer and AI-reviewer suggestions adopted much less than human ones.
Evidence, context & provenance
Evidence boundaryObserves only PRs with discussions; not controlled.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Evaluation methodology and baselines
13 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Human-AI Synergy in Agentic Code Review",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Zhong",
"source": "https://arxiv.org/abs/2603.15911",
"contribution": "Human-on-agent discussions are longer and AI-reviewer suggestions adopted much less than human ones.",
"limitation": "Observes only PRs with discussions; not controlled.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W108PolimiIn the published NDSE2 bibliography
Chen · 2025
Some model self preference reflects genuine quality, but biased acceptance of its own wrong answers persists.
Evidence, context & provenance
Evidence boundaryCode benchmark limited to MBPP+, not production PRs.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · AI as reviewer · Independent verification and oracles · Benchmarks and construct validity
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Do LLM Evaluators Prefer Themselves for a Reason?",
"kind": "paper",
"group": "Verification & review",
"year": "2025",
"lead_author": "Chen",
"source": "https://arxiv.org/abs/2504.03846",
"contribution": "Some model self preference reflects genuine quality, but biased acceptance of its own wrong answers persists.",
"limitation": "Code benchmark limited to MBPP+, not production PRs.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W109PolimiIn the published NDSE2 bibliography
McAleese · 2024
Model critics can help identify injected defects, especially combined with human scrutiny.
Evidence, context & provenance
Evidence boundaryCompared against contractors reviewing snippets without repository knowledge.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Separate reviewer effort, comment usefulness and defect detection.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. AI as reviewer · Automation bias and trust
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "LLM Critics Help Catch LLM Bugs",
"kind": "paper",
"group": "Verification & review",
"year": "2024",
"lead_author": "McAleese",
"source": "https://arxiv.org/abs/2407.00215",
"contribution": "Model critics can help identify injected defects, especially combined with human scrutiny.",
"limitation": "Compared against contractors reviewing snippets without repository knowledge.",
"manual_use": "Separate reviewer effort, comment usefulness and defect detection.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2024",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W110PolimiIn the published NDSE2 bibliography
Passi · 2022
Expert skepticism expressed in surveys need not predict actual reliance on automation.
Evidence, context & provenance
Evidence boundaryReview not focused on coding.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Automation bias and trust
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Overreliance on AI: Literature Review",
"kind": "paper",
"group": "Verification & review",
"year": "2022",
"lead_author": "Passi",
"source": "https://www.microsoft.com/en-us/research/publication/overreliance-on-ai-literature-review/",
"contribution": "Expert skepticism expressed in surveys need not predict actual reliance on automation.",
"limitation": "Review not focused on coding.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2022",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://www.microsoft.com/en-us/research/publication/overreliance-on-ai-literature-review/",
"pass7_identity_source_class": "institutional_technical_report",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W111PolimiIn the published NDSE2 bibliography
Whitfill · 2026
Automated benchmark passage overestimates practical merge-readiness for single-attempt agent patches.
Evidence, context & provenance
Evidence boundaryAgent had no reviewer feedback/CI; benchmark is not complete agent capability.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Independent verification and oracles · Benchmarks and construct validity
18 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Many SWE-bench-Passing PRs Would Not Be Merged into Main",
"kind": "paper",
"group": "Verification & review",
"year": "2026",
"lead_author": "Whitfill",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-verification-review",
"contribution": "Automated benchmark passage overestimates practical merge-readiness for single-attempt agent patches.",
"limitation": "Agent had no reviewer feedback/CI; benchmark is not complete agent capability.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH07:PRIMARY_HTML_INSPECTED; CH08:FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "3",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass4_audit_primary_url": "https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main/",
"pass4_audit_version": "METR original research note (2026-03-10)",
"pass4_audit_depth": "ORIGINAL_HTML_METHODS_AND_RESULTS_SELECTED",
"pass4_audit_methods_locators": "Data/Methods, reviewer protocol, figures 1-3, uncertainty caveat",
"pass4_audit_sample": "4 maintainers, 3 of 12 repos, 95 tasks, 296 AI patch reviews and 47 gold patches",
"pass4_audit_observation": "Automatic grader merge-decision gap 24.2 pp (SE 2.7); gold-baseline merge 68%",
"pass4_audit_boundary": "Review hypothetically in historical repos, no CI or iterative developer feedback; cluster/generalization limits",
"pass4_audit_original_data_replication": "NO",
"pass4_audit_provenance": "PASS4_SELECTED_PRIMARY_2026-10-10",
"pass4_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C07\", \"C:C08\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W112PolimiIn the published NDSE2 bibliography
Asdaque · 2026
Inexperienced agent-tool contributors impose disproportionate maintainer review effort and receive lower acceptance.
Evidence, context & provenance
Evidence boundaryExperience classification and PR selection limit causal inference.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Review capacity and maintainers · Skills formation and teaching · Dataset provenance and dependency
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Novice Developers Produce Larger Review Overhead for Project Maintainers while Vibe Coding",
"kind": "paper",
"group": "Skills & learning",
"year": "2026",
"lead_author": "Asdaque",
"source": "https://arxiv.org/abs/2602.23905",
"contribution": "Inexperienced agent-tool contributors impose disproportionate maintainer review effort and receive lower acceptance.",
"limitation": "Experience classification and PR selection limit causal inference.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W113PolimiIn the published NDSE2 bibliography
Zi · 2025
Novices cannot reliably tell whether generated programs satisfy plain-language intent.
Evidence, context & provenance
Evidence boundaryIntroductory students and short Python tasks; not professionals.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Design exercises that retain comprehension and calibrated independent judgment.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Specifications and governance · Human cognitive load · Skills formation and teaching · System understanding and rationale
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "I Would Have Written My Code Differently: Beginners Struggle to Understand LLM-Generated Code",
"kind": "paper",
"group": "Skills & learning",
"year": "2025",
"lead_author": "Zi",
"source": "https://arxiv.org/abs/2504.19037",
"contribution": "Novices cannot reliably tell whether generated programs satisfy plain-language intent.",
"limitation": "Introductory students and short Python tasks; not professionals.",
"manual_use": "Design exercises that retain comprehension and calibrated independent judgment.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:ARXIV_PDF_PAGES_VERIFIED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W114PolimiIn the published NDSE2 bibliography
Peitek · 2022
Selective code reading and cognitive engagement predict efficacy better than crude years-of-experience indicators.
Evidence, context & provenance
Evidence boundaryThirty-seven programmers; correlation not pedagogy intervention.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Design exercises that retain comprehension and calibrated independent judgment.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · Cognitive offloading and skill debt
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Correlates of programmer efficacy and their link to experience: a combined EEG and eye-tracking study",
"kind": "paper",
"group": "Skills & learning",
"year": "2022",
"lead_author": "Peitek",
"source": "https://doi.org/10.1145/3540250.3549084",
"contribution": "Selective code reading and cognitive engagement predict efficacy better than crude years-of-experience indicators.",
"limitation": "Thirty-seven programmers; correlation not pedagogy intervention.",
"manual_use": "Design exercises that retain comprehension and calibrated independent judgment.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2022",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://doi.org/10.1145/3540250.3549084",
"publication_status_note": "2022 conference version; arXiv uploaded 2023; full text pending",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PUBLISHER_DOI_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1145/3540250.3549084",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W115PolimiIn the published NDSE2 bibliography
Begel · 2008
Newly hired developers spend substantial time communicating and reconstructing team knowledge.
Evidence, context & provenance
Evidence boundaryEight graduates in one company; predates AI.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Design exercises that retain comprehension and calibrated independent judgment.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Skills formation and teaching · Employment and early-career roles
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Novice Software Developers, All Over Again",
"kind": "paper",
"group": "Skills & learning",
"year": "2008",
"lead_author": "Begel",
"source": "https://www.microsoft.com/en-us/research/publication/novice-software-developers-all-over-again/",
"contribution": "Newly hired developers spend substantial time communicating and reconstructing team knowledge.",
"limitation": "Eight graduates in one company; predates AI.",
"manual_use": "Design exercises that retain comprehension and calibrated independent judgment.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2008",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://www.microsoft.com/en-us/research/publication/novice-software-developers-all-over-again/",
"publication_status_note": "2008 developer onboarding, pre-AI context",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_INSTITUTION_PUBLICATION",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1145/1404520.1404522",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W116PolimiIn the published NDSE2 bibliography
Burelli · 2026
Physiological signals under AI assistance suggest reduced engagement despite apparent code-generation progress.
Evidence, context & provenance
Evidence boundaryStudents, two writing tasks and different unverified signal endpoints.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Design exercises that retain comprehension and calibrated independent judgment.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Human cognitive load · Skills formation and teaching · Cognitive offloading and skill debt · System understanding and rationale
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Using Biometrics to Understand AI-Assisted Coding Performance and its Perception",
"kind": "paper",
"group": "Skills & learning",
"year": "2026",
"lead_author": "Burelli",
"source": "https://arxiv.org/abs/2606.20598",
"contribution": "Physiological signals under AI assistance suggest reduced engagement despite apparent code-generation progress.",
"limitation": "Students, two writing tasks and different unverified signal endpoints.",
"manual_use": "Design exercises that retain comprehension and calibrated independent judgment.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W117PolimiIn the published NDSE2 bibliography
Caosun · 2026
Short-run assistance can undermine incentives to acquire knowledge and create delayed productivity costs.
Evidence, context & provenance
Evidence boundaryTheoretical mechanism needs longitudinal field validation.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Human cognitive load · Cognitive offloading and skill debt · Economic costs and return
14 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Augmentation Trap: AI Productivity and the Cost of Cognitive Offloading",
"kind": "paper",
"group": "Skills & learning",
"year": "2026",
"lead_author": "Caosun",
"source": "https://arxiv.org/abs/2604.03501",
"contribution": "Short-run assistance can undermine incentives to acquire knowledge and create delayed productivity costs.",
"limitation": "Theoretical mechanism needs longitudinal field validation.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/abs/2604.03501",
"publication_status_note": "Theoretical/economic model, NOT a randomized programming outcome",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PREPRINT_PRIMARY_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W118PolimiIn the published NDSE2 bibliography
Shen · 2026
Access to AI can reduce independent conceptual learning even without faster task completion.
Evidence, context & provenance
Evidence boundaryControlled learning quiz; not long-run professional performance. Immediate quiz only, not delayed retention. Source percentage normalization is unresolved; Figure 10.2 is qualitative. See Appendix B and ISSUE:V6:W118_NORMALIZATION.
Source type, as recorded: Controlled human or field experiment
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Independent verification and oracles · Skills formation and teaching · Evaluation methodology and baselines
29 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "How AI Impacts Skill Formation",
"kind": "paper",
"group": "Skills & learning",
"year": "2026",
"lead_author": "Shen",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-skills-learning",
"contribution": "Access to AI can reduce independent conceptual learning even without faster task completion.",
"limitation": "Controlled learning quiz; not long-run professional performance.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Controlled human or field experiment",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:ARXIV_V2_PDF_PAGES_VERIFIED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W119PolimiIn the published NDSE2 bibliography
Garicano · 2026
Automation of junior work may erode apprenticeship and reduce the future skilled-worker pipeline.
Evidence, context & provenance
Evidence boundaryModel predictions, not observed labor dynamics.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Design exercises that retain comprehension and calibrated independent judgment.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Skills formation and teaching · Onboarding and developer fluency · Employment and early-career roles · Remote work and apprenticeship · Economic costs and return
9 recorded source relationships
- ← W115 · Novice Software Developers, All Over Again
inferred_topic_similarity · edge e02044 · key 0 - → W120 · Automation, AI, and the Intergenerational Transmission of Knowledge
inferred_topic_similarity · edge e02115 · key 0 - → W131 · The Broken Ladder: AI, Remote Work, and Early-Career Hiring
inferred_topic_similarity · edge e02116 · key 0 - → W165 · Please Turn Your Cameras On: Remote Onboarding of Software Developers during a Pandemic
inferred_topic_similarity · edge e02117 · key 0 - → W127 · Remote Work Leaves Younger Workers Sidelined
inferred_topic_similarity · edge e02118 · key 0 - ← NDSE2:C10 · Developer Learning, Mentoring and Comprehension Debt
NDSE2_CHAPTER_CITES · edge u20261018_00891 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:EN - ← NDSE2:C10 · Developer Learning, Mentoring and Comprehension Debt
NDSE2_CHAPTER_CITES · edge u20261018_00914 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:ITA - ← NDSE2:C12 · Economics, Jobs, Risk and Evidence-Based Adoption
NDSE2_CHAPTER_CITES · edge u20261018_01086 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:EN - ← NDSE2:C12 · Economics, Jobs, Risk and Evidence-Based Adoption
NDSE2_CHAPTER_CITES · edge u20261018_01109 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:ITA
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Training in the Age of AI: A Theory of Career Viability",
"kind": "paper",
"group": "Skills & learning",
"year": "2026",
"lead_author": "Garicano",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-skills-learning",
"contribution": "Automation of junior work may erode apprenticeship and reduce the future skilled-worker pipeline.",
"limitation": "Model predictions, not observed labor dynamics.",
"manual_use": "Design exercises that retain comprehension and calibrated independent judgment.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:AUTHOR_INSTITUTIONAL_METADATA_ABSTRACT",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\", \"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W120PolimiIn the published NDSE2 bibliography
Ide · 2026
Automation changes expert-novice collaboration and may weaken practical knowledge transfer across cohorts.
Evidence, context & provenance
Evidence boundaryAbstract assumptions require organizational testing.
Source type, as recorded: Theoretical OLG task model (not empirical labor data)
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Agent failure and escalation · Skills formation and teaching · Employment and early-career roles · Remote work and apprenticeship · Economic costs and return · Evaluation methodology and baselines · Human-agent collaboration
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Automation, AI, and the Intergenerational Transmission of Knowledge",
"kind": "paper",
"group": "Skills & learning",
"year": "2026",
"lead_author": "Ide",
"source": "https://arxiv.org/abs/2507.16078",
"contribution": "Automation changes expert-novice collaboration and may weaken practical knowledge transfer across cohorts.",
"limitation": "Abstract assumptions require organizational testing.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Theoretical OLG task model (not empirical labor data)",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/abs/2507.16078",
"publication_status_note": "Theoretical model of intergenerational knowledge transmission",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PREPRINT_PRIMARY_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass2_source_locator": "https://arxiv.org/html/2507.16078",
"pass2_source_version": "arXiv HTML banner v9 2026-06-08; body 2026-08-24",
"pass2_inspection_scope": "Abstract and task-learning OLG model mechanism and appendix headings; frozen PDF-date check pending",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NOT_APPLICABLE",
"pass2_research_design": "Theoretical task-based overlapping-generations tacit knowledge diffusion",
"pass2_critical_limitation": "Conditional theoretical results not empirical skill or employment impacts; inconsistent date across HTML/banner",
"pass2_audit_date": "2026-10-10",
"pass5_audit_date": "2026-10-10",
"pass5_version": "arXiv:2507.16078v1 PDF July 2025 versus v9 PDF June 2026; same-version HTML date Aug 2026 mismatch",
"pass5_primary_locator": "https://arxiv.org/pdf/2507.16078v1 ; https://arxiv.org/pdf/2507.16078v9",
"pass5_inspection_depth": "PRIMARY_VERSIONED_PDFS_INTRO_AND_CURRENT_MODEL_PROPOSITION_4_EQUATION_11",
"pass5_page_scope": "v1 PDF p.1; v9 PDF pp.1-4 and pp.25-27 (Eq.11, Prop.4, Figure3); v9 HTML banner date checked",
"pass5_sample": "Mathematical overlapping-generation theory, no empirical software-team sample",
"pass5_main_finding": "v1 has 0.05-0.35 pp projected growth-loss scenarios; v9 derives learning composition mechanism even with full novice employment, and an immediate-positive/long-run-negative output ratio under stated parameters. Equation (11) independently recomputed at Figure3 parameters.",
"pass5_residual_limitation": "No empirical data or observed human developer retention; HTML v9 displays Aug 24 2026 although v9 PDF cover is June 9, 2026; rendered source date inconsistency remains.",
"pass5_original_data_replicated": "NO",
"pass5_provenance": "PASS5_2026-10-10_SCOPED_PRIMARY_AND_VERSIONED_ABSTRACT",
"pass6_date": "2026-10-10",
"pass6_locator": "https://arxiv.org/abs/2507.16078",
"pass6_inspection_depth": "ARXIV_HISTORY_V9_PDF_V9_EXPERIMENTAL_HTML",
"pass6_page_scope": "arxiv history line34, v9 PDF p0, HTML lines70 and 1087",
"pass6_finding": "v9 upload June 8; PDF cover June 9; experimental HTML internal August24",
"pass6_limitation": "HTML_TRANSFORMATION_DATE_CAUSE_UNKNOWN",
"pass6_original_data_replicated": "NO",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W121PolimiIn the published NDSE2 bibliography
Kosmyna · 2025
Cognitive offloading changes measures of engagement and recall, a possible analogy for programming learning.
Evidence, context & provenance
Evidence boundaryEssay writing rather than code; causality and interpretation contested.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Developer workflow and adoption · Independent verification and oracles · Human cognitive load · Skills formation and teaching · Cognitive offloading and skill debt · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task",
"kind": "paper",
"group": "Skills & learning",
"year": "2025",
"lead_author": "Kosmyna",
"source": "https://arxiv.org/abs/2506.08872",
"contribution": "Cognitive offloading changes measures of engagement and recall, a possible analogy for programming learning.",
"limitation": "Essay writing rather than code; causality and interpretation contested.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "Essay-writing EEG, not programming experiment",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/abs/2506.08872",
"publication_status_note": "Essay writing experiment, not software developer retention trial",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PREPRINT_PRIMARY_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W122PolimiIn the published NDSE2 bibliography
Osborn · 2026
AI relocates effort from syntax construction to mental models, judgment, decomposition and integration.
Evidence, context & provenance
Evidence boundaryThought piece, not a measured trend.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Design exercises that retain comprehension and calibrated independent judgment.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · System understanding and rationale · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI Didn't Make Programming Easier. It Just Made It Differently Difficult",
"kind": "paper",
"group": "Skills & learning",
"year": "2026",
"lead_author": "Osborn",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-skills-learning",
"contribution": "AI relocates effort from syntax construction to mental models, judgment, decomposition and integration.",
"limitation": "Thought piece, not a measured trend.",
"manual_use": "Design exercises that retain comprehension and calibrated independent judgment.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C01\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W123PolimiIn the published NDSE2 bibliography
Baird · 2026
Early adopter firms show higher software hiring, especially junior hiring, contradicting simple substitution stories.
Evidence, context & provenance
Evidence boundaryFirm adoption selection and platform measures complicate causal attribution.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Developer workflow and adoption · Employment and early-career roles · Evaluation methodology and baselines
17 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Firms' GitHub Copilot Adoption and Labor Market Outcomes for Software Engineers",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Baird",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"contribution": "Early adopter firms show higher software hiring, especially junior hiring, contradicting simple substitution stories.",
"limitation": "Firm adoption selection and platform measures complicate causal attribution.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W124PolimiIn the published NDSE2 bibliography
Baslandze · 2026
Leaders anticipate augmentation and work redesign, creating a comparison between aspirations and observed labor data.
Evidence, context & provenance
Evidence boundaryPlans and opinions not realized employment.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Employment and early-career roles
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Baslandze",
"source": "https://www.nber.org/papers/w34984",
"contribution": "Leaders anticipate augmentation and work redesign, creating a comparison between aspirations and observed labor data.",
"limitation": "Plans and opinions not realized employment.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://www.nber.org/papers/w34984",
"pass7_identity_source_class": "institutional_working_paper",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W125PolimiIn the published NDSE2 bibliography
Brynjolfsson · 2026
Young workers in AI-exposed occupations show relative employment declines in US administrative data.
Evidence, context & provenance
Evidence boundaryExposure and employment shifts do not uniquely identify AI causation.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Distinguish hiring composition, worker outcomes and causal attribution.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles · Evaluation methodology and baselines
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Brynjolfsson",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"contribution": "Young workers in AI-exposed occupations show relative employment declines in US administrative data.",
"limitation": "Exposure and employment shifts do not uniquely identify AI causation.",
"manual_use": "Distinguish hiring composition, worker outcomes and causal attribution.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W126PolimiIn the published NDSE2 bibliography
Chandar · 2026
AI adopters change experience requirements toward senior roles across multiple countries.
Evidence, context & provenance
Evidence boundaryVacancy wording and changes do not equal actual net employment.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "How Does AI Change Labor Demand? Evidence from 41 Countries",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Chandar",
"source": "https://digitaleconomy.stanford.edu/publication/how-does-ai-change-labor-demand/",
"contribution": "AI adopters change experience requirements toward senior roles across multiple countries.",
"limitation": "Vacancy wording and changes do not equal actual net employment.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://digitaleconomy.stanford.edu/publication/how-does-ai-change-labor-demand/",
"publication_status_note": "Cross-country labor demand; job posting and hiring are distinct from employment",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_INSTITUTION_PAPER_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Stanford Digital Economy Lab publication landing record; methods and original microdata not independently audited",
"endpoint_boundary": "Job postings/hiring and labor demand in 41 countries; neither a measure of software output nor a universal employment effect.",
"pass2_source_locator": "https://digitaleconomy.stanford.edu/app/uploads/2026/09/HowDoesAIChangeLaborDemand.pdf",
"pass2_source_version": "September 2026 primary 156-page paper",
"pass2_inspection_scope": "Abstract pp 2-3 Sections 4-5 and Figure 3 printed pp 27-28",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NO",
"pass2_research_design": "Instrumented matched foreign multinational affiliate adoption/event study",
"pass2_critical_limitation": "Junior share is not junior headcount; observed treatment via AI job-posting proxy and exclusion restrictions",
"pass2_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W127PolimiIn the published NDSE2 bibliography
Emanuel · 2026
Remote working may impair early-career access and mentorship and confound post-2022 AI hiring interpretations.
Evidence, context & provenance
Evidence boundaryRemote work and AI adoption overlap; country context matters.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Developer workflow and adoption · Agent context and persistent memory · Employment and early-career roles · Remote work and apprenticeship
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Remote Work Leaves Younger Workers Sidelined",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Emanuel",
"source": "https://libertystreeteconomics.newyorkfed.org/2026/06/remote-work-leaves-younger-workers-sidelined/",
"contribution": "Remote working may impair early-career access and mentorship and confound post-2022 AI hiring interpretations.",
"limitation": "Remote work and AI adoption overlap; country context matters.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://libertystreeteconomics.newyorkfed.org/2026/06/remote-work-leaves-younger-workers-sidelined/",
"pass7_identity_source_class": "institutional_primary_commentary",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W128PolimiIn the published NDSE2 bibliography
Facius · 2026
Early Norwegian data do not reveal the strong job losses predicted by some US exposure analyses.
Evidence, context & provenance
Evidence boundaryFollow-up window and adoption rates may differ.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Employment and early-career roles
13 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Labor Market Consequences of Generative AI: Early Evidence from Norway",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Facius",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"contribution": "Early Norwegian data do not reveal the strong job losses predicted by some US exposure analyses.",
"limitation": "Follow-up window and adoption rates may differ.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W129PolimiIn the published NDSE2 bibliography
Kauhanen · 2026
Finnish youth workforce data show no detectable AI-induced collapse in the observed horizon.
Evidence, context & provenance
Evidence boundaryNon-detection is not proof of zero long-term effects.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI Has Not Impacted the Youth Labor Market in Finland",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Kauhanen",
"source": "https://www.etla.fi/en/publications/working-papers-en/ai-has-not-impacted-the-youth-labor-market-in-finland/",
"contribution": "Finnish youth workforce data show no detectable AI-induced collapse in the observed horizon.",
"limitation": "Non-detection is not proof of zero long-term effects.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://www.etla.fi/en/publications/working-papers-en/ai-has-not-impacted-the-youth-labor-market-in-finland/",
"publication_status_note": "Finnish labor evidence cannot automatically generalize internationally",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_INSTITUTION_PAPER_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"pass2_source_locator": "https://www.etla.fi/wp-content/uploads/ETLA-Working-Papers-135.pdf",
"pass2_source_version": "ETLA Working Papers 135 January 2026",
"pass2_inspection_scope": "Original figures 3 7 8 and methods and conclusions printed pp 10 15 16 18",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NO",
"pass2_research_design": "Registry headcounts and exposure quantile descriptive plus firm fixed effect contrasts",
"pass2_critical_limitation": "No controlled experiment assigning chatbot use; negative detectable exposure gradient not proof no AI effect",
"pass2_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W130PolimiIn the published NDSE2 bibliography
Kharazian · 2026
High AI spending correlates with growing firm employment including entry hires.
Evidence, context & provenance
Evidence boundarySurviving growth firms and spending are endogenous.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles · Evaluation methodology and baselines
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Kharazian",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"contribution": "High AI spending correlates with growing firm employment including entry hires.",
"limitation": "Surviving growth firms and spending are endogenous.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W131PolimiIn the published NDSE2 bibliography
Lambert · 2026
Junior hiring contraction may be more closely explained by remote-work effects than AI deployment alone.
Evidence, context & provenance
Evidence boundaryAttribution depends on exposure and comparator specification.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Specifications and governance · Employment and early-career roles · Remote work and apprenticeship
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Broken Ladder: AI, Remote Work, and Early-Career Hiring",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Lambert",
"source": "https://wrap.warwick.ac.uk/id/eprint/202349/",
"contribution": "Junior hiring contraction may be more closely explained by remote-work effects than AI deployment alone.",
"limitation": "Attribution depends on exposure and comparator specification.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_pass2_section_scoped",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://wrap.warwick.ac.uk/id/eprint/202349/",
"publication_status_note": "Junior hiring and job ads; controls for WFH alter significance; confirm actual version before quoting numbers",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_INSTITUTION_PAPER_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"pass2_source_locator": "https://cep.lse.ac.uk/pubs/download/dp2193.pdf",
"pass2_source_version": "LSE CEP Discussion Paper 2193 June 2026 primary 83-page PDF",
"pass2_inspection_scope": "Abstract plus original identification printed around pp 12-13 and joint treatment caution",
"pass2_methods_status": "PRIMARY_MANUSCRIPT_METHODS_AND_TABLES_SELECTED",
"pass2_replication_status": "NO",
"pass2_research_design": "Observational continuous-exposure difference in differences with WFH co-treatment controls",
"pass2_critical_limitation": "AI and WFH exposure rank correlation 0.77 and strong parallel exposure trend conditions; remote work not randomized",
"pass2_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W132PolimiIn the published NDSE2 bibliography
Skoglund · 2026
Historical occupational turnover suggests task change and adaptation alongside displacement risk.
Evidence, context & provenance
Evidence boundaryLong historical analogies may not hold at current AI speed.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Vulnerability and secure code · Ownership and knowledge transfer · Employment and early-career roles · Long-running agent execution
4 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "New Jobs in 140 Years of Data: Why the AI Displacement Fear Is Overstated — and What to Worry About Instead",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Skoglund",
"source": "https://cepr.org/voxeu/columns/new-jobs-140-years-data-why-ai-displacement-fear-overstated-and-what-worry-about",
"contribution": "Historical occupational turnover suggests task change and adaptation alongside displacement risk.",
"limitation": "Long historical analogies may not hold at current AI speed.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://cepr.org/voxeu/columns/new-jobs-140-years-data-why-ai-displacement-fear-overstated-and-what-worry-about",
"pass7_identity_source_class": "author_primary_analysis",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W133PolimiIn the published NDSE2 bibliography
Yotzov · 2026
Most companies currently report no major AI-driven employment changes, while future staffing expectations modestly decline.
Evidence, context & provenance
Evidence boundarySelf-report and delayed effects; broad not SWE-specific.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Distinguish hiring composition, worker outcomes and causal attribution.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Firm Data on AI",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Yotzov",
"source": "https://www.nber.org/papers/w34836",
"contribution": "Most companies currently report no major AI-driven employment changes, while future staffing expectations modestly decline.",
"limitation": "Self-report and delayed effects; broad not SWE-specific.",
"manual_use": "Distinguish hiring composition, worker outcomes and causal attribution.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://www.nber.org/papers/w34836",
"pass7_identity_source_class": "institutional_working_paper",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W134PolimiIn the published NDSE2 bibliography
Hampole · 2025
AI-exposed roles can see pay pressure even when headcount and occupation effects differ.
Evidence, context & provenance
Evidence boundaryReported wage effect excludes SW engineering roles, limiting manual applicability.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles
15 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Artificial Intelligence and the Labor Market",
"kind": "paper",
"group": "Jobs & roles",
"year": "2025",
"lead_author": "Hampole",
"source": "https://www.nber.org/papers/w33509",
"contribution": "AI-exposed roles can see pay pressure even when headcount and occupation effects differ.",
"limitation": "Reported wage effect excludes SW engineering roles, limiting manual applicability.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://www.nber.org/papers/w33509",
"pass7_identity_source_class": "institutional_working_paper",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W135PolimiIn the published NDSE2 bibliography
Hosseini · 2025
AI adoption may shift demand toward experienced software workers at the expense of junior vacancies.
Evidence, context & provenance
Evidence boundaryJob-posting measurement and selection bias.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Human cognitive load · Employment and early-career roles · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data",
"kind": "paper",
"group": "Jobs & roles",
"year": "2025",
"lead_author": "Hosseini",
"source": "https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555",
"contribution": "AI adoption may shift demand toward experienced software workers at the expense of junior vacancies.",
"limitation": "Job-posting measurement and selection bias.",
"manual_use": "Use to build a reviewer-calibration exercise with injected wrong suggestions and measured detection rates.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555",
"pass7_identity_source_class": "primary_versioned_manuscript",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W136PolimiIn the published NDSE2 bibliography
Humlum · 2025
No material early change in worker earnings or hours despite adoption; workplace tasks can shift first.
Evidence, context & provenance
Evidence boundaryEarly interval, labor market protections and country context.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Agent context and persistent memory · Employment and early-career roles
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI",
"kind": "paper",
"group": "Jobs & roles",
"year": "2025",
"lead_author": "Humlum",
"source": "https://www.nber.org/papers/w33777",
"contribution": "No material early change in worker earnings or hours despite adoption; workplace tasks can shift first.",
"limitation": "Early interval, labor market protections and country context.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_institutional_record_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://www.nber.org/papers/w33777",
"publication_status_note": "Denmark administrative earnings/hours; not universally comparable to hiring postings",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "WORKING_PAPER_INSTITUTION",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"pass2_source_locator": "https://www.nber.org/papers/w33777",
"pass2_source_version": "NBER WP33777 original May 2025 and institutional March 2026 revision",
"pass2_inspection_scope": "Original institutional publication summary for revision; full numerical tables not independently inspected",
"pass2_methods_status": "PRIMARY_INSTITUTIONAL_RECORD_ONLY",
"pass2_replication_status": "NO",
"pass2_research_design": "Linked Danish chatbot-use surveys and wage/hour employment records",
"pass2_critical_limitation": "Full PDF coefficient scale and identification not checked; cannot infer universal or long-run zero impact",
"pass2_audit_date": "2026-10-10",
"pass5_audit_date": "2026-10-10",
"pass5_version": "RFBerlin Discussion Paper 078/26, original PDF 13 March 2026; NBER revision March 2026",
"pass5_primary_locator": "https://www.rfberlin.com/wp-content/uploads/2026/03/26078.pdf",
"pass5_inspection_depth": "ORIGINAL_132_PAGE_PDF_DESIGN_EQUATIONS_FIGURES_AND_SELECTED_RESULTS",
"pass5_page_scope": "PDF pp.3-7, 15-16, 20-26, 29-31; printed pp.1-5, 13-14, 18-24, 27-29",
"pass5_sample": "Approx 25,000 workers / 7,000 workplaces, 11 exposed occupations, linked to administrative Danish data through Dec 2024",
"pass5_main_finding": "Eq(3) OLS pooled DiD with worker-specific linear trends and workplace-occupation clustered SE; Figure 5 dynamic CIs exclude >3% earnings shifts; pooled Figure 7 excludes >2%; survey adoption/descriptive differences should not be causalized.",
"pass5_residual_limitation": "Primary tables beyond selected figures not exhaustively inspected; Statistics Denmark linked raw data unavailable for independent regression.",
"pass5_original_data_replicated": "NO",
"pass5_provenance": "PASS5_2026-10-10_SCOPED_PRIMARY_AND_VERSIONED_ABSTRACT",
"pass6_date": "2026-10-10",
"pass6_locator": "https://www.rfberlin.com/wp-content/uploads/2026/03/26078.pdf",
"pass6_inspection_depth": "PRIMARY_132_PAGE_PDF_METHODS_AND_FIGURE_TEXT",
"pass6_page_scope": "PDF p24-25; printed p22-23",
"pass6_finding": "Eq2 dynamic/eq3 pooled; 9/4% pre-adoption earnings premiums; 3/2% different confidence intervals",
"pass6_limitation": "LICENSED_DENMARK_MICRODATA_UNAVAILABLE",
"pass6_original_data_replicated": "NO",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W137PolimiIn the published NDSE2 bibliography
Aidala · 2026
Posted development vacancies disproportionately target seniority compared with entry-level roles.
Evidence, context & provenance
Evidence boundaryDescriptive share changes cannot attribute cause to AI.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Labor Market Is Tilting Toward Seniority",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Aidala",
"source": "https://hiringlab.indeed.com/2026/07/23/the-labor-market-is-tilting-toward-seniority/",
"contribution": "Posted development vacancies disproportionately target seniority compared with entry-level roles.",
"limitation": "Descriptive share changes cannot attribute cause to AI.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://hiringlab.indeed.com/2026/07/23/the-labor-market-is-tilting-toward-seniority/",
"pass7_identity_source_class": "producer_primary_report",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W138PolimiIn the published NDSE2 bibliography
Athreya · 2026
Employers may adjust planned recruiting before adjusting existing headcount.
Evidence, context & provenance
Evidence boundaryReports intent and sentiment rather than realized effects.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Specifications and governance · Employment and early-career roles
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI's Impact on Labor and Hiring",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Athreya",
"source": "https://libertystreeteconomics.newyorkfed.org/2026/08/ais-impact-on-labor-and-hiring/",
"contribution": "Employers may adjust planned recruiting before adjusting existing headcount.",
"limitation": "Reports intent and sentiment rather than realized effects.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://libertystreeteconomics.newyorkfed.org/2026/08/ais-impact-on-labor-and-hiring/",
"pass7_identity_source_class": "institutional_primary_commentary",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W139PolimiIn the published NDSE2 bibliography
Audoly · 2026
Software vacancy decline predates widespread ChatGPT use in some US data, weakening simplistic causal narratives.
Evidence, context & provenance
Evidence boundaryVacancy data are incomplete and other economic shocks matter.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Employment and early-career roles · Economic costs and return
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Do Job Postings Show Early Labor-Market Effects of AI?",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Audoly",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"contribution": "Software vacancy decline predates widespread ChatGPT use in some US data, weakening simplistic causal narratives.",
"limitation": "Vacancy data are incomplete and other economic shocks matter.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W140PolimiIn the published NDSE2 bibliography
Gallacher · 2026
Some developer vacancies rise after new coding agent capabilities become available.
Evidence, context & provenance
Evidence boundaryProduct release timing is not randomized AI exposure.
Source type, as recorded: Observational job-postings trend analysis (not a randomized field experiment)
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Employment and early-career roles
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI and Job Postings: From Destruction to Creation?",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Gallacher",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"contribution": "Some developer vacancies rise after new coding agent capabilities become available.",
"limitation": "Product release timing is not randomized AI exposure.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Controlled human or field experiment",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Controlled human or field experiment",
"reviewed_evidence_class": "Observational job-postings trend analysis (not a randomized field experiment)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "2026",
"publication_source_url": "https://hiringlab.indeed.com/2026/07/08/ai-and-job-postings-from-destruction-to-creation/",
"publication_status_note": "Indeed Hiring Lab observational commentary",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W141PolimiIn the published NDSE2 bibliography
Iscenko · 2026
Maps adoption by tasks; tool usage does not directly imply task replacement or labor demand change.
Evidence, context & provenance
Evidence boundaryOne provider and incomplete agent usage coverage.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Employment and early-career roles
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Iscenko",
"source": "https://arxiv.org/abs/2608.00038",
"contribution": "Maps adoption by tasks; tool usage does not directly imply task replacement or labor demand change.",
"limitation": "One provider and incomplete agent usage coverage.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://arxiv.org/abs/2608.00038",
"pass7_identity_source_class": "author_preprint",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W142PolimiIn the published NDSE2 bibliography
Nunn · 2026
Alternative US estimation fails to identify significant net AI-driven labor weakening.
Evidence, context & provenance
Evidence boundaryMethod and observation window could miss subgroup impacts.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Agent failure and escalation · Employment and early-career roles · Evaluation methodology and baselines
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AI Is Probably Not (Yet) the Reason for Labor Market Weakening",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "Nunn",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"contribution": "Alternative US estimation fails to identify significant net AI-driven labor weakening.",
"limitation": "Method and observation window could miss subgroup impacts.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W143PolimiIn the published NDSE2 bibliography
Kinder · 2025
Aggregate labor indicators remain more stable than narratives of immediate massive AI displacement.
Evidence, context & provenance
Evidence boundaryAggregate trends can mask cohort, wage and industry effects.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Employment and early-career roles · Evaluation methodology and baselines
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "New data show no AI jobs apocalypse — for now",
"kind": "paper",
"group": "Jobs & roles",
"year": "2025",
"lead_author": "Kinder",
"source": "https://www.brookings.edu/articles/new-data-show-no-ai-jobs-apocalypse-for-now/",
"contribution": "Aggregate labor indicators remain more stable than narratives of immediate massive AI displacement.",
"limitation": "Aggregate trends can mask cohort, wage and industry effects.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://www.brookings.edu/articles/new-data-show-no-ai-jobs-apocalypse-for-now/",
"pass7_identity_source_class": "institutional_primary_commentary",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W144PolimiIn the published NDSE2 bibliography
PwC · 2026
Rapidly changing skill demands and experience requirements may reshape developer career ladders.
Evidence, context & provenance
Evidence boundaryVendor interpretation and task-exposure proxies; not causal.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Skills formation and teaching · Employment and early-career roles
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "2026 Global AI Jobs Barometer: Two futures for jobs in an AI era",
"kind": "paper",
"group": "Jobs & roles",
"year": "2026",
"lead_author": "PwC",
"source": "https://www.pwc.com/gx/en/issues/artificial-intelligence/publications/artificial-intelligence-study.html",
"contribution": "Rapidly changing skill demands and experience requirements may reshape developer career ladders.",
"limitation": "Vendor interpretation and task-exposure proxies; not causal.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://www.pwc.com/gx/en/issues/artificial-intelligence/publications/artificial-intelligence-study.html",
"pass7_identity_source_class": "producer_primary_report",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-jobs-roles",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W145PolimiIn the published NDSE2 bibliography
DORA · 2026
Offers a cross-functional return-on-investment framework for AI-assisted delivery beyond raw code productivity.
Evidence, context & provenance
Evidence boundaryROI 39% and eight-month payback are scenarios/model outputs, not measured experimental financial returns.
Source type, as recorded: Practitioner guidance / ROI calculator (not an experimental study)
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Human cognitive load · Economic costs and return · Evaluation methodology and baselines
21 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "ROI of AI-Assisted Software Development",
"kind": "paper",
"group": "Economics",
"year": "2026",
"lead_author": "DORA",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-economics",
"contribution": "Offers a cross-functional return-on-investment framework for AI-assisted delivery beyond raw code productivity.",
"limitation": "ROI 39% and eight-month payback are scenarios/model outputs, not measured experimental financial returns.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Practitioner guidance / ROI calculator (not an experimental study)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "2026",
"publication_source_url": "https://dora.dev/ai/roi/report/",
"publication_status_note": "DORA practitioner ROI framework/report",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C12\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W146PolimiIn the published NDSE2 bibliography
Chui · 2026
Positions legacy system context and secure deployment as adoption constraints, and stresses both acceleration and uneven results.
Evidence, context & provenance
Evidence boundarySurvey opinions and mixed trend indicators are not causal SWE productivity measures.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Controlled productivity evidence · Developer workflow and adoption · Vulnerability and secure code · System understanding and rationale · Agent context and persistent memory · Evaluation methodology and baselines
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "McKinsey Technology Trends Outlook 2026",
"kind": "paper",
"group": "Economics",
"year": "2026",
"lead_author": "Chui",
"source": "https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-top-trends-in-tech",
"contribution": "Positions legacy system context and secure deployment as adoption constraints, and stresses both acceleration and uneven results.",
"limitation": "Survey opinions and mixed trend indicators are not causal SWE productivity measures.",
"manual_use": "Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-top-trends-in-tech",
"pass7_identity_source_class": "producer_primary_report",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-economics",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W147PolimiIn the published NDSE2 bibliography
Yang · 2026
Projects adopt distinct governance stances to regulate low-cost agent PR supply and scarce maintainer review capacity.
Evidence, context & provenance
Evidence boundaryStrategy taxonomy does not establish effectiveness.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Specifications and governance · Economic costs and return · Human-agent collaboration
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "To Ban or not to Ban? How Open Source Projects Govern GenAI Contributions",
"kind": "paper",
"group": "Engineering structure",
"year": "2026",
"lead_author": "Yang",
"source": "https://arxiv.org/abs/2603.26487",
"contribution": "Projects adopt distinct governance stances to regulate low-cost agent PR supply and scarce maintainer review capacity.",
"limitation": "Strategy taxonomy does not establish effectiveness.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH11:ORIGINAL_PDF_METHODS_AND_TABLES",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W148PolimiIn the published NDSE2 bibliography
Lin · 2026
Process discipline can improve generated code on selected functional tests compared with unconstrained runs.
Evidence, context & provenance
Evidence boundarySmall benchmark and GPT-3.5-era setup; process emulation not production governance.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Benchmarks and construct validity · Architecture and modularity · Specifications and governance
16 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "SOEN-101: Code Generation by Emulating Software Process Models Using Large Language Model Agents",
"kind": "paper",
"group": "Engineering structure",
"year": "2026",
"lead_author": "Lin",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-engineering-structure",
"contribution": "Process discipline can improve generated code on selected functional tests compared with unconstrained runs.",
"limitation": "Small benchmark and GPT-3.5-era setup; process emulation not production governance.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "HumanEval/MBPP coding benchmarks, not live agile field study",
"publication_year_from_independent_source": "2025",
"publication_source_url": "https://doi.org/10.1109/ICSE55347.2025.00140",
"publication_status_note": "ICSE 2025 (arXiv manuscript first posted 2024)",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH07:PUBLISHER_ABSTRACT_ONLY; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass4_audit_primary_url": "https://arxiv.org/pdf/2403.15852",
"pass4_audit_version": "Original 14-page arXiv:2403.15852v2 author manuscript, 2024-10-31; 2025 IEEE conference record separately checked",
"pass4_audit_depth": "AUTHOR_PREPRINT_PDF_METHODS_TABLES_II_III_PAGE_INSPECTED",
"pass4_audit_methods_locators": "Author manuscript PDF pp. 6-7, Table II and Table III, experiment GPT-3.5 and five repeats; benchmark MBPP-582 test anomaly",
"pass4_audit_sample": "164 HumanEval problems; 427 MBPP problems; HumanEval-ET/MBPP-ET extended test sets; five GPT-3.5 runs with T=0.8",
"pass4_audit_observation": "Mean pass@1 Scrum 75.2/65.5/82.5/56.7; Table II baseline 64.4/49.8/77.5/53.9; MBPP-582 mismatched test input shown in Listing 1",
"pass4_audit_boundary": "Final ICSE PDF tables not compared with author v2; reported p-values not recomputed from run-level outputs; no human team rollout evidence",
"pass4_audit_original_data_replication": "NO",
"pass4_audit_provenance": "PASS4_SELECTED_PRIMARY_2026-10-10",
"pass4_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C04\", \"C:C07\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W149PolimiIn the published NDSE2 bibliography
Chen · 2026
Explicit project policies correlate with fewer measured code smells and richer human PR discussion without suppressing AI use.
Evidence, context & provenance
Evidence boundaryEight-week matched window cannot rule out proactive-maintainer selection.
Source type, as recorded: Quasi-experimental / matched observation
Editorial use. Make policy, responsibility and bounded workflow executable.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Technical debt and complexity · Specifications and governance · Evaluation methodology and baselines · Human-agent collaboration
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Making AI Visible, Not Vanished: How AI Policies Reshape Developer Experience on GitHub",
"kind": "paper",
"group": "Engineering structure",
"year": "2026",
"lead_author": "Chen",
"source": "https://arxiv.org/abs/2608.03329",
"contribution": "Explicit project policies correlate with fewer measured code smells and richer human PR discussion without suppressing AI use.",
"limitation": "Eight-week matched window cannot rule out proactive-maintainer selection.",
"manual_use": "Make policy, responsibility and bounded workflow executable.",
"evidence_class": "Quasi-experimental / matched observation",
"editorial_priority": "A",
"bibliography_year": "2026",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "Quasi-experimental / matched observation",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_PENDING; CH11:ORIGINAL_PDF_METHODS_AND_TABLES",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W150PolimiIn the published NDSE2 bibliography
Hora · 2026
Most policies allow AI while specifying disclosure and human responsibility, suggesting workable rather than absolutist governance.
Evidence, context & provenance
Evidence boundaryClassifies rules but does not measure their safety effect.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make policy, responsibility and bounded workflow executable.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Specifications and governance · Evaluation methodology and baselines · Human-agent collaboration
33 recorded source relationships
- ← W043 · We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs
inferred_topic_similarity · edge e00852 · key 0 - ← W077 · Evaluating and Mitigating the Misguidance Effect of Buggy Code in LLM-Generated Unit Tests
inferred_topic_similarity · edge e01412 · key 0 - ← W147 · To Ban or not to Ban? How Open Source Projects Govern GenAI Contributions
governance taxonomy and prevalence · edge e02547 · key 0 - ← W147 · To Ban or not to Ban? How Open Source Projects Govern GenAI Contributions
inferred_topic_similarity · edge e02548 · key 1 - ← W149 · Making AI Visible, Not Vanished: How AI Policies Reshape Developer Experience on GitHub
observational evaluation of practice · edge e02595 · key 0 - ← W149 · Making AI Visible, Not Vanished: How AI Policies Reshape Developer Experience on GitHub
inferred_topic_similarity · edge e02596 · key 1 - ← O041 · Leading Effective Engineering Teams in the Age of GenAI
EDITORIAL_COMPARISON_NOT_CITATION · edge e04661 · key 0 - ← B010 · The Adaptation Imperative---What Companies Must Do Now
ND_EDITORIAL_COMPARISON · edge x06375 · key ND-b88a8d6811dcf0c0 - ← B016 · PART IV: Organizational Foundation
ND_EDITORIAL_COMPARISON · edge x06508 · key ND-98dd2e9008783e5c - ← B025 · From Single Agent to Multi-Agent Workflows
ND_EDITORIAL_COMPARISON · edge x06696 · key ND-c035ae053cc8461e - ← B027 · Team Collaboration \& Knowledge Sharing
ND_EDITORIAL_COMPARISON · edge x06746 · key ND-3d474d3e3fc647cb - ← B028 · Security, Compliance \& Governance
ND_EDITORIAL_COMPARISON · edge x06768 · key ND-7ace4aebcd134263 - ← B038 · The Shift
ND_EDITORIAL_COMPARISON · edge x07018 · key ND-0e718466cdb637c0 - ← B041 · The Integration
ND_EDITORIAL_COMPARISON · edge x07078 · key ND-e5e2dcb9d6d47dea - ← B049 · The Shift
ND_EDITORIAL_COMPARISON · edge x07284 · key ND-30cdd7358d37292a - ← B058 · Design Boundaries: What to Adopt, What to Avoid
ND_EDITORIAL_COMPARISON · edge x07551 · key ND-9e2e2b2f5e227c78 - ← B059 · Constitution and Architecture: The Framework Foundation
ND_EDITORIAL_COMPARISON · edge x07576 · key ND-261c099744b0ed6a - ← B060 · The Agent Pipeline and Gates
ND_EDITORIAL_COMPARISON · edge x07617 · key ND-f7f53fc7285fd9af - ← B062 · Autonomous Execution: Modes, Budgets, and Governance
ND_EDITORIAL_COMPARISON · edge x07707 · key ND-db21d03a57f85723 - ← B065 · How to Design a New Framework
ND_EDITORIAL_COMPARISON · edge x07790 · key ND-c2849a856340c96a - ← B068 · Template Reference
ND_EDITORIAL_COMPARISON · edge x07859 · key ND-93fe1344edef46e9 - ← B070 · Enterprise SDD
ND_EDITORIAL_COMPARISON · edge x07916 · key ND-48fb604497d419bb - ← B094 · Scheduling and Overnight Execution
ND_EDITORIAL_COMPARISON · edge x08477 · key ND-54af9408342f9893 - ← B095 · Multi-Agent Orchestration
ND_EDITORIAL_COMPARISON · edge x08502 · key ND-abf451b1afda131a - ← B103 · Summary
ND_EDITORIAL_COMPARISON · edge x08661 · key ND-fd04fce82a178091 - ← B104 · Non-Deterministic Loop Engineering: A Critical Assessment
ND_EDITORIAL_COMPARISON · edge x08697 · key ND-0f089bc37909ab61 - ← B105 · The Organization Becomes a Loop
ND_EDITORIAL_COMPARISON · edge x08725 · key ND-724210992effed06 - ← NDSE2:C04 · Intentions, Contracts and Iterative Change
NDSE2_CHAPTER_CITES · edge u20261018_00325 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:EN - ← NDSE2:C04 · Intentions, Contracts and Iterative Change
NDSE2_CHAPTER_CITES · edge u20261018_00340 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:ITA - ← NDSE2:C07 · Independent Verification and the Oracle Problem
NDSE2_CHAPTER_CITES · edge u20261018_00575 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:EN - ← NDSE2:C07 · Independent Verification and the Oracle Problem
NDSE2_CHAPTER_CITES · edge u20261018_00594 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:ITA - ← NDSE2:C11 · Teams, Governance and the Engineering Organization
NDSE2_CHAPTER_CITES · edge u20261018_00996 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:EN - ← NDSE2:C11 · Teams, Governance and the Engineering Organization
NDSE2_CHAPTER_CITES · edge u20261018_01015 · key UPDATE_20261018:NDSE2_CHAPTER_CITES:ITA
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "We Permit the Use of AI, but [...]: The Landscape of AI Policies in Popular Open Source Projects",
"kind": "paper",
"group": "Engineering structure",
"year": "2026",
"lead_author": "Hora",
"source": "https://arxiv.org/abs/2609.07542",
"contribution": "Most policies allow AI while specifying disclosure and human responsibility, suggesting workable rather than absolutist governance.",
"limitation": "Classifies rules but does not measure their safety effect.",
"manual_use": "Make policy, responsibility and bounded workflow executable.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH07:FULLTEXT_HTML; CH11:ORIGINAL_PDF_METHODS_AND_TABLES",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass4_audit_primary_url": "https://arxiv.org/html/2609.07542",
"pass4_audit_version": "arXiv:2609.07542v1 (2026-09-07)",
"pass4_audit_depth": "ORIGINAL_HTML_METHODS_AND_TABLES_SELECTED",
"pass4_audit_methods_locators": "Methods Sections 2.2-2.4; RQ1 Table 1; RQ2 Fig 4; limitations",
"pass4_audit_sample": "281 purposively selected policies (245/2000 popular repos + 36 added); 92 dedicated policy files",
"pass4_audit_observation": "150/281 mention disclosure obligation; 137/281 require disclosure; 180 double-coded labels, 162 agree",
"pass4_audit_boundary": "Selection and coder agreement do not measure policy effectiveness; 118-policy earlier sample not independent",
"pass4_audit_original_data_replication": "NO",
"pass4_audit_provenance": "PASS4_SELECTED_PRIMARY_2026-10-10",
"pass4_audit_date": "2026-10-10",
"pass9_scope": "original HTML cited Zenodo DOI located, unadjusted coder agreement, no source-record download",
"ndse2_actual_chapters_json": "[\"C:C04\", \"C:C07\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W151PolimiIn the published NDSE2 bibliography
Qiu · 2026
Predefined action blueprints can constrain agent behavior and outperform unrestricted planning on tested workflows.
Evidence, context & provenance
Evidence boundaryDomain not typical SWE repositories; benchmark-specific.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Independent verification and oracles · Benchmarks and construct validity · Architecture and modularity · Specifications and governance · Long-running agent execution
18 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Blueprint First, Model Second: A Framework for Deterministic LLM Workflow",
"kind": "paper",
"group": "Engineering structure",
"year": "2026",
"lead_author": "Qiu",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-engineering-structure",
"contribution": "Predefined action blueprints can constrain agent behavior and outperform unrestricted planning on tested workflows.",
"limitation": "Domain not typical SWE repositories; benchmark-specific.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "Tool/workflow tau-bench, not real coding-team outcomes",
"publication_year_from_independent_source": "2025",
"publication_source_url": "https://arxiv.org/abs/2508.02721",
"publication_status_note": "arXiv preprint (first posted August 2025)",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH07:FULLTEXT_HTML; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "3",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass3_audit_primary_url": "https://arxiv.org/html/2508.02721v2",
"pass3_audit_version": "arXiv:2508.02721v2 (2026-06-16)",
"pass3_audit_depth": "ORIGINAL_HTML_SELECTED_METHODS_AND_TABLES",
"pass3_audit_methods_locators": "IV Table I; RQ2 hard constraints; experimental settings",
"pass3_audit_sample": "TravelPlanner 1000 tasks; Claude Sonnet 4; three seeds",
"pass3_audit_observation": "Final pass 35.56% SCA versus ATLAS 18.00%; hard-constraint macro 38.89 versus 42.00",
"pass3_audit_boundary": "Hard constraint outcomes trade off; benchmark does not validate real policy authority",
"pass3_audit_original_data_replication": "NO",
"pass3_audit_provenance": "PASS3_PRINTED_EDITION_RECONSTRUCTED_2026-10-10",
"pass3_audit_date": "2026-10-10",
"pass9_scope": "v1/v2 original HTML Table 1/I-II; McNemar from continuous mean not verifiable",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C04\", \"C:C07\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W152PolimiIn the published NDSE2 bibliography
Hassan · 2026
Provides an architecture for organizing agentic SWE research around dependable processes, collaboration and evaluation.
Evidence, context & provenance
Evidence boundaryRoadmap is not evidence of achieved agent reliability.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Architecture and modularity · Employment and early-career roles · Human-agent collaboration
35 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Agentic Software Engineering: Foundational Pillars and a Research Roadmap",
"kind": "paper",
"group": "Engineering structure",
"year": "2026",
"lead_author": "Hassan",
"source": "https://arxiv.org/abs/2509.06216",
"contribution": "Provides an architecture for organizing agentic SWE research around dependable processes, collaboration and evaluation.",
"limitation": "Roadmap is not evidence of achieved agent reliability.",
"manual_use": "Present as one estimand within a multi-country labor evidence table; avoid deterministic jobs predictions.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/abs/2509.06216",
"publication_status_note": "Roadmap; initial submission 2025, metadata may reflect 2026 revision",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PREPRINT_PRIMARY_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "3",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass4_audit_primary_url": "https://arxiv.org/html/2509.06216",
"pass4_audit_version": "arXiv:2509.06216v3 (2026-06-24)",
"pass4_audit_depth": "ORIGINAL_HTML_CONCEPTUAL_METHOD",
"pass4_audit_methods_locators": "Sections 1,4-7 and named proposed artifacts",
"pass4_audit_sample": "No intervention cohort; authored roadmap SASE",
"pass4_audit_observation": "Four pillars Actors/Processes/Tools/Artifacts; SE4H/SE4A duality; proposed scripts",
"pass4_audit_boundary": "Design proposal not validated impact on multi-agent team performance",
"pass4_audit_original_data_replication": "NO",
"pass4_audit_provenance": "PASS4_SELECTED_PRIMARY_2026-10-10",
"pass4_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C04\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W153PolimiIn the published NDSE2 bibliography
Hoda · 2026
Expands the scope of agentic SWE to requirements, design, teamwork and organization-wide responsibility.
Evidence, context & provenance
Evidence boundaryFraming proposal; no new controlled outcome evidence.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Evaluation methodology and baselines
13 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Toward Agentic Software Engineering Beyond Code: Framing Vision, Values, and Vocabulary",
"kind": "paper",
"group": "Engineering structure",
"year": "2026",
"lead_author": "Hoda",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-engineering-structure",
"contribution": "Expands the scope of agentic SWE to requirements, design, teamwork and organization-wide responsibility.",
"limitation": "Framing proposal; no new controlled outcome evidence.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass4_audit_primary_url": "https://arxiv.org/html/2510.19692",
"pass4_audit_version": "arXiv:2510.19692v2 (2026-02-17)",
"pass4_audit_depth": "ORIGINAL_HTML_CONCEPTUAL_METHOD",
"pass4_audit_methods_locators": "Abstract, values and vocabulary sections",
"pass4_audit_sample": "No original randomized or natural-experiment sample",
"pass4_audit_observation": "Advocates lifecycle-wide agentic SE and preliminary CRAFT values",
"pass4_audit_boundary": "Normative framework not empirical proof of organizational effects",
"pass4_audit_original_data_replication": "NO",
"pass4_audit_provenance": "PASS4_SELECTED_PRIMARY_2026-10-10",
"pass4_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C04\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W154PolimiIn the published NDSE2 bibliography
Ahmad · 2026
Delegated code generation may leave developers with latent understanding deficits analogous to accumulated technical debt.
Evidence, context & provenance
Evidence boundaryThematic self-reports from students, without effect-size or objective comprehension baseline.
Source type, as recorded: Qualitative student diary study (not commercial self-report)
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Technical debt and complexity · Human cognitive load · Skills formation and teaching · Cognitive offloading and skill debt · System understanding and rationale · Evaluation methodology and baselines
19 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Comprehension Debt in GenAI-Assisted Software Engineering Projects",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2026",
"lead_author": "Ahmad",
"source": "https://arxiv.org/abs/2604.13277",
"contribution": "Delegated code generation may leave developers with latent understanding deficits analogous to accumulated technical debt.",
"limitation": "Thematic self-reports from students, without effect-size or objective comprehension baseline.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Qualitative student diary study (not commercial self-report)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "2026",
"publication_source_url": "https://arxiv.org/abs/2604.13277",
"publication_status_note": "Student diary qualitative study (arXiv/EASE 2026)",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING; CH10:ARXIV_PDF_PAGES_VERIFIED",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W155PolimiIn the published NDSE2 bibliography
Anderson · 2026
ChatGPT may aid completion of changes while developers still devote most time to reconstructing existing-system behavior.
Evidence, context & provenance
Evidence boundaryTen advanced students; results exploratory and not statistically tested.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Skills formation and teaching · System understanding and rationale
15 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "ChatGPT: Friend or Foe When Comprehending and Changing Unfamiliar Code",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2026",
"lead_author": "Anderson",
"source": "https://arxiv.org/abs/2605.10702",
"contribution": "ChatGPT may aid completion of changes while developers still devote most time to reconstructing existing-system behavior.",
"limitation": "Ten advanced students; results exploratory and not statistically tested.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING; CH10:ARXIV_PDF_PAGES_VERIFIED",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C09\", \"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W156PolimiIn the published NDSE2 bibliography
Kara · 2026
AI-assisted tours curated by experts outperform AI-generated-only code tours on a short comprehension test.
Evidence, context & provenance
Evidence boundaryFive participants; no no-AI arm or long-run onboarding outcome.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Human cognitive load · Skills formation and teaching · Onboarding and developer fluency · System understanding and rationale
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "LACY: Simulating Expert Mentoring for Software Onboarding with Code Tours",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2026",
"lead_author": "Kara",
"source": "https://arxiv.org/abs/2603.25391",
"contribution": "AI-assisted tours curated by experts outperform AI-generated-only code tours on a short comprehension test.",
"limitation": "Five participants; no no-AI arm or long-run onboarding outcome.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING; CH10:ARXIV_PDF_PAGES_VERIFIED_PUBLISHER_METADATA_CONFIRMED",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W157PolimiIn the published NDSE2 bibliography
Rafique · 2026
Typed state pages and enforced write policy protect essential agent memory against compaction and reset failures.
Evidence, context & provenance
Evidence boundaryGuarantees conditional on token-budget working set; limited trace and synthetic evaluation.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Agent failure and escalation · Dependencies and agent supply chain · Specifications and governance · Agent context and persistent memory · Long-running agent execution · Human-agent collaboration · Dataset provenance and dependency
27 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "ClawVM: Harness-Managed Virtual Memory for Stateful Tool-Using LLM Agents",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2026",
"lead_author": "Rafique",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"contribution": "Typed state pages and enforced write policy protect essential agent memory against compaction and reset failures.",
"limitation": "Guarantees conditional on token-budget working set; limited trace and synthetic evaluation.",
"manual_use": "Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:ORIGINAL_PDF_VERIFIED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C09\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W158PolimiIn the published NDSE2 bibliography
Tie · 2026
Wrong replies, overload and context loss can cause users to abandon AI assistance in a demanding web-development task.
Evidence, context & provenance
Evidence boundaryChat assistant rather than a modern autonomous tool-using agent.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Agent failure and escalation · Agent context and persistent memory · Long-running agent execution
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Should I Give Up Now? Investigating LLM Pitfalls in Software Engineering",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2026",
"lead_author": "Tie",
"source": "https://arxiv.org/abs/2411.09916",
"contribution": "Wrong replies, overload and context loss can cause users to abandon AI assistance in a demanding web-development task.",
"limitation": "Chat assistant rather than a modern autonomous tool-using agent.",
"manual_use": "Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C09\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W159PolimiIn the published NDSE2 bibliography
Xu · 2026
Hybrid episodic and semantic storage with retention policies can improve long-horizon consistency and token efficiency.
Evidence, context & provenance
Evidence boundarySynthetic low-code trials; missing real SWE longitudinal validation.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Agent failure and escalation · Agent context and persistent memory · Economic costs and return · Long-running agent execution · Human-agent collaboration
23 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Memory Management and Contextual Consistency for Long-Running Low-Code Agents",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2026",
"lead_author": "Xu",
"source": "https://arxiv.org/abs/2509.25250",
"contribution": "Hybrid episodic and semantic storage with retention policies can improve long-horizon consistency and token efficiency.",
"limitation": "Synthetic low-code trials; missing real SWE longitudinal validation.",
"manual_use": "Translate into a context-handoff pattern separating working context, durable facts and verified system knowledge.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/abs/2509.25250",
"publication_status_note": "Low-code agent-memory study; no software organization field results",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PREPRINT_PRIMARY_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C09\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W160PolimiIn the published NDSE2 bibliography
Anokhin · 2025
Combining semantic facts with ordered episodes improves decisions over unstructured history in text environments.
Evidence, context & provenance
Evidence boundaryEvaluated in games, not a production software repo.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Benchmarks and construct validity · Skills formation and teaching · Agent context and persistent memory · Long-running agent execution
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2025",
"lead_author": "Anokhin",
"source": "https://arxiv.org/abs/2407.04363",
"contribution": "Combining semantic facts with ordered episodes improves decisions over unstructured history in text environments.",
"limitation": "Evaluated in games, not a production software repo.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "General agent episodic knowledge graphs, not code-maintenance intervention",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PUBLISHED_PDF_VERIFIED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C09\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W161PolimiIn the published NDSE2 bibliography
Koana · 2024
Ownership-defect correlations vary across datasets and cannot be used as immutable quality laws.
Evidence, context & provenance
Evidence boundaryNine-project replication changes sign relative to earlier Microsoft studies.
Source type, as recorded: Literature synthesis
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Ownership and knowledge transfer · Evaluation methodology and baselines · Dataset provenance and dependency
13 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Examining ownership models in software teams: A systematic literature review and a replication study",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2024",
"lead_author": "Koana",
"source": "https://doi.org/10.1007/s10664-024-10538-5",
"contribution": "Ownership-defect correlations vary across datasets and cannot be used as immutable quality laws.",
"limitation": "Nine-project replication changes sign relative to earlier Microsoft studies.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Literature synthesis",
"editorial_priority": "B",
"bibliography_year": "2024",
"original_evidence_class": "Literature synthesis",
"reviewed_evidence_class": "Literature synthesis",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1007/s10664-024-10538-5",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W162PolimiIn the published NDSE2 bibliography
Wyrich · 2024
Program-comprehension experiments rarely test code personally authored by participants, leaving a major knowledge-transfer confound.
Evidence, context & provenance
Evidence boundaryMaps pre-AI experiments; does not test agents directly.
Source type, as recorded: SYSTEMATIC_MAPPING_95_PRE_AI_EXPERIMENTS
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Human cognitive load · System understanding and rationale
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "40 Years of Designing Code Comprehension Experiments: A Systematic Mapping Study",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2024",
"lead_author": "Wyrich",
"source": "https://doi.org/10.1145/3626522",
"contribution": "Program-comprehension experiments rarely test code personally authored by participants, leaving a major knowledge-transfer confound.",
"limitation": "Maps pre-AI experiments; does not test agents directly.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "SYSTEMATIC_MAPPING_95_PRE_AI_EXPERIMENTS",
"editorial_priority": "B",
"bibliography_year": "2023",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "SYSTEMATIC_MAPPING_95_PRE_AI_EXPERIMENTS",
"full_manuscript_audited": "yes_selected_methods",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "2024",
"publication_source_url": "https://doi.org/10.1145/3626522",
"publication_status_note": "ACM Computing Surveys 56(4), 2024; 95 studies; 2026 corrigendum DOI 10.1145/3838282",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "FULLTEXT_PREPRINT_AND_PUBLISHED_CORRIGENDUM",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_MANUSCRIPT_SECTIONS",
"audit_locator": "Original author preprint 2206.11102 Sections 2-5, PDF pp. 1-43; publication 2024; 2026 corrigendum 10.1145/3838282, section 1",
"doi": "10.1145/3626522",
"correction_warning": "2026 corrigendum DOI 10.1145/3838282: primary studies S1–S95 formerly mismatched; use corrected IDs only.",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W163PolimiIn the published NDSE2 bibliography
Ju · 2021
Onboarding requires orchestrated starter tasks, mentoring and access to team knowledge rather than merely fast initial commits.
Evidence, context & provenance
Evidence boundaryMainly qualitative; duration not measured.
Source type, as recorded: MIXED_METHOD_SINGLE_COMPANY_CASE_STUDY
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Skills formation and teaching · Onboarding and developer fluency · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "A Case Study of Onboarding in Software Teams: Tasks and Strategies",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2021",
"lead_author": "Ju",
"source": "https://homepages.dcc.ufmg.br/~figueiredo/disciplinas/papers/icse2021ju.pdf",
"contribution": "Onboarding requires orchestrated starter tasks, mentoring and access to team knowledge rather than merely fast initial commits.",
"limitation": "Mainly qualitative; duration not measured.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "MIXED_METHOD_SINGLE_COMPANY_CASE_STUDY",
"editorial_priority": "C",
"bibliography_year": "2021",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "MIXED_METHOD_SINGLE_COMPANY_CASE_STUDY",
"full_manuscript_audited": "yes_selected_methods",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://homepages.dcc.ufmg.br/~figueiredo/disciplinas/papers/icse2021ju.pdf",
"publication_status_note": "32 developer and 15 manager interviews; 189 developer/37 manager survey; Microsoft only; case study not random assignment",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "FULLTEXT_PRIMARY_METHODS_AND_RESULTS",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_MANUSCRIPT_SECTIONS",
"audit_locator": "ICSE 2021 PDF pp. 1-3 (abstract, Methods, sample recruitment), Section VI threats",
"doi": "10.1109/ICSE43902.2021.00063",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W164PolimiIn the published NDSE2 bibliography
Peitek · 2021
Common complexity metrics capture only part of human cognitive effort; text size and vocabulary also influence comprehension.
Evidence, context & provenance
Evidence boundaryNineteen participants, artificial snippets.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Technical debt and complexity · Human cognitive load · System understanding and rationale
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Program Comprehension and Code Complexity Metrics: An fMRI Study",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2021",
"lead_author": "Peitek",
"source": "https://github.com/brains-on-code/fMRI-complexity-metrics-icse2021",
"contribution": "Common complexity metrics capture only part of human cognitive effort; text size and vocabulary also influence comprehension.",
"limitation": "Nineteen participants, artificial snippets.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2021",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://github.com/brains-on-code/fMRI-complexity-metrics-icse2021",
"pass7_identity_source_class": "author_software_repository",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W165PolimiIn the published NDSE2 bibliography
Rodeghero · 2021
Remote onboarding puts coordination and first-PR progress into focus; early contribution is not complete project fluency.
Evidence, context & provenance
Evidence boundarySurvey report; does not measure months to mastery.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Onboarding and developer fluency · Remote work and apprenticeship · Evaluation methodology and baselines
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Please Turn Your Cameras On: Remote Onboarding of Software Developers during a Pandemic",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2021",
"lead_author": "Rodeghero",
"source": "https://arxiv.org/abs/2011.08130",
"contribution": "Remote onboarding puts coordination and first-PR progress into focus; early contribution is not complete project fluency.",
"limitation": "Survey report; does not measure months to mastery.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2021",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W166PolimiIn the published NDSE2 bibliography
Ivanova · 2020
Code comprehension draws heavily on general executive cognition, not ordinary natural-language comprehension alone.
Evidence, context & provenance
Evidence boundaryPsychology of small programs rather than codebase ownership.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · System understanding and rationale · Ownership and knowledge transfer
7 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Comprehension of computer code relies primarily on domain-general executive brain regions",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2020",
"lead_author": "Ivanova",
"source": "https://doi.org/10.7554/eLife.58906",
"contribution": "Code comprehension draws heavily on general executive cognition, not ordinary natural-language comprehension alone.",
"limitation": "Psychology of small programs rather than codebase ownership.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2020",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.7554/eLife.58906",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W167PolimiIn the published NDSE2 bibliography
Peitek · 2020
Programming-language familiarity lowers cognitive effort during code reading.
Evidence, context & provenance
Evidence boundaryDoes not distinguish own code from unfamiliar team code.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · System understanding and rationale
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "A Look into Programmers' Heads",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2020",
"lead_author": "Peitek",
"source": "https://doi.org/10.1109/TSE.2018.2863303",
"contribution": "Programming-language familiarity lowers cognitive effort during code reading.",
"limitation": "Does not distinguish own code from unfamiliar team code.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2020",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1109/TSE.2018.2863303",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W168PolimiIn the published NDSE2 bibliography
Krüger · 2018
File-specific knowledge decays after contributions; own edits and share of code correlate with perceived familiarity.
Evidence, context & provenance
Evidence boundarySelf-reported familiarity is not validated recall.
Source type, as recorded: Self-report / commercial evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Automation bias and trust
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Do you remember this source code?",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2018",
"lead_author": "Krüger",
"source": "https://doi.org/10.1145/3180155.3180215",
"contribution": "File-specific knowledge decays after contributions; own edits and share of code correlate with perceived familiarity.",
"limitation": "Self-reported familiarity is not validated recall.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Self-report / commercial evidence",
"editorial_priority": "C",
"bibliography_year": "2018",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "Self-report / commercial evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://doi.org/10.1145/3180155.3180215",
"publication_status_note": "60 open-source developer survey, not an AI memory validation",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PUBLISHER_DOI_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1145/3180155.3180215",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C09\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W169PolimiIn the published NDSE2 bibliography
Peitek · 2018
Brain/gaze instrumentation for code understanding has important signal and measurement constraints.
Evidence, context & provenance
Evidence boundaryMore about measurement feasibility than AI intervention.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · System understanding and rationale · Evaluation methodology and baselines
14 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Simultaneous Measurement of Program Comprehension with fMRI and Eye Tracking: A Case Study",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2018",
"lead_author": "Peitek",
"source": "https://github.com/brains-on-code/simultaneous-fmri-and-eyetracking",
"contribution": "Brain/gaze instrumentation for code understanding has important signal and measurement constraints.",
"limitation": "More about measurement feasibility than AI intervention.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2018",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://github.com/brains-on-code/simultaneous-fmri-and-eyetracking",
"pass7_identity_source_class": "author_software_repository",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W170PolimiIn the published NDSE2 bibliography
Xia · 2018
Most development effort is spent understanding and navigating existing code, especially by newcomers.
Evidence, context & provenance
Evidence boundaryUnderstanding inferred from telemetry, not observed; authorship unseparated.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · Onboarding and developer fluency · System understanding and rationale · Ownership and knowledge transfer · Evaluation methodology and baselines
16 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Measuring Program Comprehension: A Large-Scale Field Study with Professionals",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2018",
"lead_author": "Xia",
"source": "https://xin-xia.github.io/publication/TSE17.pdf",
"contribution": "Most development effort is spent understanding and navigating existing code, especially by newcomers.",
"limitation": "Understanding inferred from telemetry, not observed; authorship unseparated.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2018",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PUBLISHER_ABSTRACT_VERIFIED; CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass4_audit_primary_url": "https://baolingfeng.github.io/papers/tsecomprehension.pdf",
"pass4_audit_version": "Author-hosted 23-page PDF versus TSE 44(10) 2018 publisher record",
"pass4_audit_depth": "ORIGINAL_PDF_SELECTED_TABLES_JOURNAL_METADATA",
"pass4_audit_methods_locators": "Author PDF p1, p7 Table3, p9 Table5, p22 threats; institutional published abstract",
"pass4_audit_sample": "Author draft 79 developers/3244 hours; published journal abstract 78/3148, seven projects",
"pass4_audit_observation": "Draft Table5 57.62% comprehension; journal abstract about 58%; draft Table3 2-developer validation",
"pass4_audit_boundary": "Version sample-count conflict remains uncorrected without published full journal tables; not AI-specific",
"pass4_audit_original_data_replication": "NO",
"pass4_audit_provenance": "PASS4_SELECTED_PRIMARY_2026-10-10",
"pass4_audit_date": "2026-10-10",
"pass6_date": "2026-10-10",
"pass6_locator": "https://baolingfeng.github.io/papers/tsecomprehension.pdf",
"pass6_inspection_depth": "AUTHOR_23_PAGE_PDF_AND_PUBLISHED_ABSTRACT",
"pass6_page_scope": "author p0; Table1 p3; Table5 p8; Monash journal metadata",
"pass6_finding": "79 and 3244 author vs 78 and 3148 journal; author Table1 sums 79, exactly one developer and 96 hours difference",
"pass6_limitation": "FINAL_JOURNAL_NUMERICAL_TABLES_UNAVAILABLE",
"pass6_original_data_replicated": "NO",
"pass8_primary_final_published_url": "https://xin-xia.github.io/publication/TSE17.pdf",
"pass8_methods": "IEEE final p954 confirms explicit 96-hour excluded participant; journal Table1 p954 still sums 79; Table5 p959 same prior rates",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C09\", \"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W171PolimiIn the published NDSE2 bibliography
Nassif · 2017
Knowledge abandoned when contributors depart is concentrated in a tail of long-unowned files.
Evidence, context & provenance
Evidence boundaryCode authorship serves as proxy for knowledge; agents threaten assumption.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Agent failure and escalation · Ownership and knowledge transfer · Evaluation methodology and baselines · Dataset provenance and dependency
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Revisiting Turnover-Induced Knowledge Loss in Software Projects",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2017",
"lead_author": "Nassif",
"source": "https://doi.org/10.1109/ICSME.2017.64",
"contribution": "Knowledge abandoned when contributors depart is concentrated in a tail of long-unowned files.",
"limitation": "Code authorship serves as proxy for knowledge; agents threaten assumption.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2017",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1109/ICSME.2017.64",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W172PolimiIn the published NDSE2 bibliography
Siegmund · 2017
Semantic cues like informative naming reduce cognitive effort during code comprehension.
Evidence, context & provenance
Evidence boundarySmall trained sample; no AI and no realistic full projects.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · System understanding and rationale · Agent context and persistent memory · Economic costs and return · Evaluation methodology and baselines · Dataset provenance and dependency
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Measuring neural efficiency of program comprehension",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2017",
"lead_author": "Siegmund",
"source": "https://doi.org/10.1145/3106237.3106268",
"contribution": "Semantic cues like informative naming reduce cognitive effort during code comprehension.",
"limitation": "Small trained sample; no AI and no realistic full projects.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2017",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1145/3106237.3106268",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W173PolimiIn the published NDSE2 bibliography
Rigby · 2016
Developer departure leaves knowledge coverage gaps; successor assignment can partially reduce them.
Evidence, context & provenance
Evidence boundaryCode-blame ownership proxy may fail when agents commit code.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Vulnerability and secure code · Ownership and knowledge transfer
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Quantifying and Mitigating Turnover-Induced Knowledge Loss: Case Studies of Chrome and a Project at Avaya",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2016",
"lead_author": "Rigby",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"contribution": "Developer departure leaves knowledge coverage gaps; successor assignment can partially reduce them.",
"limitation": "Code-blame ownership proxy may fail when agents commit code.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2016",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:ORIGINAL_PDF_VERIFIED; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C09\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W174PolimiIn the published NDSE2 bibliography
Foucault · 2015
External contributor churn can affect downstream quality and code stewardship.
Evidence, context & provenance
Evidence boundaryEffect sizes not retrieved in annotated bibliography.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Ownership and knowledge transfer
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Impact of Developer Turnover on Quality in Open-Source Software",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2015",
"lead_author": "Foucault",
"source": "https://doi.org/10.1145/2786805.2786870",
"contribution": "External contributor churn can affect downstream quality and code stewardship.",
"limitation": "Effect sizes not retrieved in annotated bibliography.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2015",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1145/2786805.2786870",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W175PolimiIn the published NDSE2 bibliography
Greiler · 2015
Weak code ownership predicts defects in some settings while intentional shared ownership can be healthy.
Evidence, context & provenance
Evidence boundaryRegression only, ownership meaning depends on organizational context.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Specifications and governance · Ownership and knowledge transfer · Agent context and persistent memory · Evaluation methodology and baselines · Dataset provenance and dependency
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Code Ownership and Software Quality: A Replication Study",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2015",
"lead_author": "Greiler",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"contribution": "Weak code ownership predicts defects in some settings while intentional shared ownership can be healthy.",
"limitation": "Regression only, ownership meaning depends on organizational context.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2015",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:AUTHOR_CAMERA_READY_PDF_LOCATED; CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C05\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W176PolimiIn the published NDSE2 bibliography
Minelli · 2015
Understanding consumes a larger share of developer time than typing in an observed Smalltalk setting.
Evidence, context & provenance
Evidence boundaryOne IDE ecosystem and mixed student/professional cohort.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Skills formation and teaching · System understanding and rationale · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "I Know What You Did Last Summer — An Investigation of How Developers Spend Their Time",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2015",
"lead_author": "Minelli",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"contribution": "Understanding consumes a larger share of developer time than typing in an observed Smalltalk setting.",
"limitation": "One IDE ecosystem and mixed student/professional cohort.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2015",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C01\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W177PolimiIn the published NDSE2 bibliography
Müller · 2015
Sensing subjective strain and flow might complement throughput metrics when evaluating human-agent collaboration.
Evidence, context & provenance
Evidence boundaryTiny sample and predictor errors; monitoring raises privacy concerns.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Developer workflow and adoption · Employment and early-career roles · Human-agent collaboration · Dataset provenance and dependency
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Stuck and Frustrated or in Flow and Happy: Sensing Developers' Emotions and Progress",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2015",
"lead_author": "Müller",
"source": "https://doi.org/10.1109/ICSE.2015.334",
"contribution": "Sensing subjective strain and flow might complement throughput metrics when evaluating human-agent collaboration.",
"limitation": "Tiny sample and predictor errors; monitoring raises privacy concerns.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2015",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"pass7_verified_primary_locator": "https://doi.org/10.1109/ICSE.2015.334",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W178PolimiIn the published NDSE2 bibliography
Rastogi · 2015
First commit arrives well before full file-level breadth and ramp-up; different activity measures plateau at different times.
Evidence, context & provenance
Evidence boundaryAbsolute durations withheld; first PR is invalid proxy for competency.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Onboarding and developer fluency · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Ramp-Up Journey of New Hires: Tug of War of Aids and Impediments",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2015",
"lead_author": "Rastogi",
"source": "https://doi.org/10.1109/ESEM.2015.7321212",
"contribution": "First commit arrives well before full file-level breadth and ramp-up; different activity measures plateau at different times.",
"limitation": "Absolute durations withheld; first PR is invalid proxy for competency.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2015",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://doi.org/10.1109/ESEM.2015.7321212",
"publication_status_note": "2015 ESEM ramp-up paper; distinct from later similarly named paper",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PUBLISHER_DOI_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1109/ESEM.2015.7321212",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W179PolimiIn the published NDSE2 bibliography
Sharafi · 2015
Existing gaze-based studies seldom isolate effects of own-code versus other-code understanding.
Evidence, context & provenance
Evidence boundaryPredates LLM tooling; study design gaps not outcome evidence.
Source type, as recorded: Literature synthesis
Editorial use. Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Developer workflow and adoption · Human cognitive load · System understanding and rationale
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "A systematic literature review on the usage of eye-tracking in software engineering",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2015",
"lead_author": "Sharafi",
"source": "https://doi.org/10.1016/j.infsof.2015.06.008",
"contribution": "Existing gaze-based studies seldom isolate effects of own-code versus other-code understanding.",
"limitation": "Predates LLM tooling; study design gaps not outcome evidence.",
"manual_use": "Compare human-on-human, human-on-agent and AI-on-human review and record review effort, precision and real defects.",
"evidence_class": "Literature synthesis",
"editorial_priority": "B",
"bibliography_year": "2015",
"original_evidence_class": "Literature synthesis",
"reviewed_evidence_class": "Literature synthesis",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1016/j.infsof.2015.06.008",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W180PolimiIn the published NDSE2 bibliography
Fritz · 2014
Multi-sensor measurements may detect task difficulty and cognitive burden more directly than edit counts.
Evidence, context & provenance
Evidence boundaryPerformance on new programmers weaker; implementation costs and privacy limits.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · Economic costs and return · Evaluation methodology and baselines
6 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Using psycho-physiological measures to assess task difficulty in software development",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2014",
"lead_author": "Fritz",
"source": "https://www.microsoft.com/en-us/research/publication/using-psycho-physiological-measures-to-assess-task-difficulty-in-software-development/",
"contribution": "Multi-sensor measurements may detect task difficulty and cognitive burden more directly than edit counts.",
"limitation": "Performance on new programmers weaker; implementation costs and privacy limits.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2014",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://www.microsoft.com/en-us/research/publication/using-psycho-physiological-measures-to-assess-task-difficulty-in-software-development/",
"pass7_identity_source_class": "institutional_publication_record",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W181PolimiIn the published NDSE2 bibliography
Siegmund · 2014
Brain activation differs between true program comprehension and superficial syntax spotting.
Evidence, context & provenance
Evidence boundaryContrived snippets with obscured identifiers.
Source type, as recorded: Benchmark / controlled laboratory
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Human cognitive load · System understanding and rationale
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Understanding understanding source code with functional magnetic resonance imaging",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2014",
"lead_author": "Siegmund",
"source": "https://doi.org/10.1145/2568225.2568252",
"contribution": "Brain activation differs between true program comprehension and superficial syntax spotting.",
"limitation": "Contrived snippets with obscured identifiers.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Benchmark / controlled laboratory",
"editorial_priority": "B",
"bibliography_year": "2014",
"original_evidence_class": "Benchmark / controlled laboratory",
"reviewed_evidence_class": "Benchmark / controlled laboratory",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://doi.org/10.1145/2568225.2568252",
"pass7_identity_source_class": "publisher_doi",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W182PolimiIn the published NDSE2 bibliography
Bird · 2011
Low-concentration human code ownership correlates with defects; suggests stewardship matters.
Evidence, context & provenance
Evidence boundaryProduct-level correlation and human-authorship assumption not portable unchanged to agent output.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Ownership and knowledge transfer
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Don't Touch My Code! Examining the Effects of Ownership on Software Quality",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2011",
"lead_author": "Bird",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"contribution": "Low-concentration human code ownership correlates with defects; suggests stewardship matters.",
"limitation": "Product-level correlation and human-authorship assumption not portable unchanged to agent output.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2011",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:PUBLISHER_RECORD_AND_AUTHOR_SOURCE_LOCATED; CH09:ORIGINAL_PDF_VERIFIED",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C05\", \"C:C09\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W183PolimiIn the published NDSE2 bibliography
Dagenais · 2010
New developers need exploration and environment setup before completing useful changes, regardless of code generation rate.
Evidence, context & provenance
Evidence boundarySmall heterogeneous IBM samples, legacy context.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Onboarding and developer fluency · System understanding and rationale · Agent context and persistent memory · Dataset provenance and dependency
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Moving into a New Software Project Landscape",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2010",
"lead_author": "Dagenais",
"source": "https://research.ibm.com/publications/moving-into-a-new-software-project-landscape",
"contribution": "New developers need exploration and environment setup before completing useful changes, regardless of code generation rate.",
"limitation": "Small heterogeneous IBM samples, legacy context.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2010",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://research.ibm.com/publications/moving-into-a-new-software-project-landscape",
"pass7_identity_source_class": "institutional_publication_record",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W184PolimiIn the published NDSE2 bibliography
Zhou · 2010
Newcomers can match task counts long before developing fluency for difficult central components.
Evidence, context & provenance
Evidence boundaryThree-plus-year mastery is a hypothesis, not directly measured universal threshold.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Onboarding and developer fluency · Evaluation methodology and baselines
10 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Developer Fluency: Achieving True Mastery in Software Projects",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2010",
"lead_author": "Zhou",
"source": "https://doi.org/10.1145/1882291.1882313",
"contribution": "Newcomers can match task counts long before developing fluency for difficult central components.",
"limitation": "Three-plus-year mastery is a hypothesis, not directly measured universal threshold.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2010",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://doi.org/10.1145/1882291.1882313",
"publication_status_note": "2010 conference; full method tables pending",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "PUBLISHER_DOI_RECORD",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1145/1882291.1882313",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W185PolimiIn the published NDSE2 bibliography
Mockus · 2009
Ownership transfer across teams or geography may lower successor effectiveness without deliberate knowledge handoff.
Evidence, context & provenance
Evidence boundaryPast corporate development model; not agent-mediated work.
Source type, as recorded: INDUSTRIAL_OBSERVATIONAL_SUCCESSION_REGRESSION
Editorial use. Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Ownership and knowledge transfer · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Succession: Measuring Transfer of Code and Developer Productivity",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2009",
"lead_author": "Mockus",
"source": "https://mockus.org/papers/succession.pdf",
"contribution": "Ownership transfer across teams or geography may lower successor effectiveness without deliberate knowledge handoff.",
"limitation": "Past corporate development model; not agent-mediated work.",
"manual_use": "Use to teach a metrics hierarchy: generated code → integrated change → deployed outcome → customer value.",
"evidence_class": "INDUSTRIAL_OBSERVATIONAL_SUCCESSION_REGRESSION",
"editorial_priority": "A",
"bibliography_year": "2009",
"original_evidence_class": "Quasi-experimental / matched observation",
"reviewed_evidence_class": "INDUSTRIAL_OBSERVATIONAL_SUCCESSION_REGRESSION",
"full_manuscript_audited": "yes_selected_methods",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://mockus.org/papers/succession.pdf",
"publication_status_note": "ICSE 2009 Sections 5 and 6, Tables 1-2; 10 identified succession cases and 1,012 mentor-follower pairs",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "FULLTEXT_PRIMARY_METHODS_TABLES",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_MANUSCRIPT_SECTIONS",
"audit_locator": "ICSE 2009 PDF p. 6 Table 1; p. 9 Table 2 (zero-based PDF page indices 5 and 8), Section 6.3-6.4",
"doi": "10.1109/ICSE.2009.5070509",
"previous_chapter_audit_registers": "CH09:ORIGINAL_ABSTRACT_AND_METADATA_VERIFIED; CH10:FULL_TEXT_REVIEW_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "3",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C09\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W186PolimiIn the published NDSE2 bibliography
LaToza · 2006
Developers struggle particularly with others' code and hidden design rationale, making explanation provenance an asset.
Evidence, context & provenance
Evidence boundaryPerceived problem ratings, not timed comprehension measures.
Source type, as recorded: EXPLORATORY_SINGLE_COMPANY_SURVEY_AND_INTERVIEWS
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Automation bias and trust · Human cognitive load · System understanding and rationale · Evaluation methodology and baselines
9 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Maintaining Mental Models: A Study of Developer Work Habits",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2006",
"lead_author": "LaToza",
"source": "https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/p492-latoza.pdf",
"contribution": "Developers struggle particularly with others' code and hidden design rationale, making explanation provenance an asset.",
"limitation": "Perceived problem ratings, not timed comprehension measures.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "EXPLORATORY_SINGLE_COMPANY_SURVEY_AND_INTERVIEWS",
"editorial_priority": "C",
"bibliography_year": "2006",
"original_evidence_class": "Self-report / commercial evidence",
"reviewed_evidence_class": "EXPLORATORY_SINGLE_COMPANY_SURVEY_AND_INTERVIEWS",
"full_manuscript_audited": "yes_selected_methods",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/p492-latoza.pdf",
"publication_status_note": "ICSE 2006 Section 3.3, 157 and 187 responses from separate invitations of 1000 each; 11 interviews; selection bias not measured",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "FULLTEXT_PRIMARY_METHODS_TABLES",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_MANUSCRIPT_SECTIONS",
"audit_locator": "ICSE 2006 PDF pp. 2-4, Section 3 method, 3.3 sampling and Table 1",
"doi": "10.1145/1134285.1134355",
"previous_chapter_audit_registers": "CH09:AUTHOR_INSTITUTIONAL_ABSTRACT_VERIFIED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C01\", \"C:C09\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W187PolimiIn the published NDSE2 bibliography
Mockus · 2001
New engineers took more than a year on average to reach a change-count plateau in one telecommunications system.
Evidence, context & provenance
Evidence boundaryOne team and counting commits; does not measure full mastery or AI effects.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Evaluation methodology and baselines
12 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Globalization by Chunking: A Quantitative Approach",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2001",
"lead_author": "Mockus",
"source": "https://mockus.org/papers/chunking.pdf",
"contribution": "New engineers took more than a year on average to reach a change-count plateau in one telecommunications system.",
"limitation": "One team and counting commits; does not measure full mastery or AI effects.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2001",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no_primary_locator_only",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://mockus.org/papers/chunking.pdf",
"publication_status_note": "Original author-hosted journal paper; full-text inspection not completed",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"source_verification_method": "AUTHOR_FULL_TEXT_LOCATOR_ONLY",
"audit_date": "2026-10-10",
"audit_depth": "PRIMARY_METADATA_LOCATOR",
"audit_locator": "Primary publication record; page-level study audit pending",
"doi": "10.1109/52.914737",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C10\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W188PolimiIn the published NDSE2 bibliography
Sim · 1998
Environmental access and mentoring dominate early onboarding, with independence possible before comprehensive system understanding.
Evidence, context & provenance
Evidence boundaryN=4; expert estimate for six-to-twelve-month productivity not measured.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Output versus shipped value · Human cognitive load · Skills formation and teaching · Onboarding and developer fluency · System understanding and rationale · Evaluation methodology and baselines
8 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Ramp-Up Problem in Software Projects: A Case Study of How Software Immigrants Naturalize",
"kind": "paper",
"group": "System knowledge & memory",
"year": "1998",
"lead_author": "Sim",
"source": "https://dl.acm.org/doi/10.5555/302163.302199",
"contribution": "Environmental access and mentoring dominate early onboarding, with independence possible before comprehensive system understanding.",
"limitation": "N=4; expert estimate for six-to-twelve-month productivity not measured.",
"manual_use": "Use as a baseline when designing AI-supported onboarding with independent comprehension and retention tests.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "1998",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH10:FULL_TEXT_REVIEW_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass7_verified_primary_locator": "https://dl.acm.org/doi/10.5555/302163.302199",
"pass7_identity_source_class": "publisher_record",
"pass7_methods_replicated": "NO",
"legacy_editorial_source": "https://www.sepolimi.com/p/se-ai-bibliography-system-knowledge-memory",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W189PolimiIn the published NDSE2 bibliography
Wheeler · 2026
Authorship-based truck-factor and ownership metrics may become misleading once agents author code without human understanding.
Evidence, context & provenance
Evidence boundaryCentral insight is conceptual; no empirical recalibration proposed.
Source type, as recorded: Theory / argument (not a new result)
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Skills formation and teaching · System understanding and rationale · Ownership and knowledge transfer
14 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "The Substrate Collapse: AI Code Generation Invalidates Authorship-Based Knowledge Metrics",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2026",
"lead_author": "Wheeler",
"source": "https://arxiv.org/abs/2606.20882",
"contribution": "Authorship-based truck-factor and ownership metrics may become misleading once agents author code without human understanding.",
"limitation": "Central insight is conceptual; no empirical recalibration proposed.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Theory / argument (not a new result)",
"editorial_priority": "C",
"bibliography_year": "2026",
"original_evidence_class": "Theory / argument (not a new result)",
"reviewed_evidence_class": "Theory / argument (not a new result)",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH05:FULL_TEXT_OPENED_METHODS_PENDING; CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING; CH11:FULL_TEXT_PENDING",
"previous_chapter_audit_count": "3",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"pass4_audit_primary_url": "https://arxiv.org/html/2606.20882",
"pass4_audit_version": "arXiv:2606.20882v1 (2026-06-18)",
"pass4_audit_depth": "ORIGINAL_HTML_CONCEPTUAL_METHOD",
"pass4_audit_methods_locators": "Premise, measurement critique and explicit scope sections",
"pass4_audit_sample": "No new developer sample or fresh authorship-truck-factor validation",
"pass4_audit_observation": "Argues commits and authorship are no longer reliable proxies for comprehension under generation",
"pass4_audit_boundary": "Argument is plausible but does not statistically prove invalidity for every repository",
"pass4_audit_original_data_replication": "NO",
"pass4_audit_provenance": "PASS4_SELECTED_PRIMARY_2026-10-10",
"pass4_audit_date": "2026-10-10",
"ndse2_actual_chapters_json": "[\"C:C05\", \"C:C09\", \"C:C11\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W190PolimiIn the published NDSE2 bibliography
Abdelsalam · 2025
Model perplexity and human confusion can correlate on difficult source lines, suggesting possible code-comprehension diagnostics.
Evidence, context & provenance
Evidence boundaryReused small EEG dataset and one model; correlation not general ability.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Preserve human system comprehension, provenance and durable agent context.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Code reuse and duplication · Human cognitive load · System understanding and rationale · Dataset provenance and dependency
5 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "How do Humans and LLMs Process Confusing Code?",
"kind": "paper",
"group": "System knowledge & memory",
"year": "2025",
"lead_author": "Abdelsalam",
"source": "https://arxiv.org/abs/2508.18547",
"contribution": "Model perplexity and human confusion can correlate on difficult source lines, suggesting possible code-comprehension diagnostics.",
"limitation": "Reused small EEG dataset and one model; correlation not general ability.",
"manual_use": "Preserve human system comprehension, provenance and durable agent context.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2025",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH09:PRIMARY_FULL_TEXT_AUDIT_PENDING",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W191PolimiIn the published NDSE2 bibliography
Ehsani · 2026
Rejected agent PRs reveal recurring limits in requirement understanding, repository fit and coordination with maintainers.
Evidence, context & provenance
Evidence boundaryFailure coding depends on visible GitHub information; shared AIDev sample.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Review capacity and maintainers · Agent failure and escalation · System understanding and rationale · Dataset provenance and dependency
34 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "Where Do AI Coding Agents Fail? An Empirical Study of Failed Agentic Pull Requests in GitHub",
"kind": "paper",
"group": "What agents can do",
"year": "2026",
"lead_author": "Ehsani",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-what-agents-can-do",
"contribution": "Rejected agent PRs reveal recurring limits in requirement understanding, repository fit and coordination with maintainers.",
"limitation": "Failure coding depends on visible GitHub information; shared AIDev sample.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:ARXIV_ABSTRACT_VERIFIED",
"previous_chapter_audit_count": "1",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C04\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W192PolimiIn the published NDSE2 bibliography
Li · 2026
Supplies a reusable 932791-PR universe and a curated 33596-PR subset that underpins many agent outcome studies.
Evidence, context & provenance
Evidence boundaryShared data are not independent replications; five agents and snapshot through mid-2025.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Independent verification and oracles · Evaluation methodology and baselines · Dataset provenance and dependency
13 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "AIDev: Studying AI Coding Agents on GitHub",
"kind": "paper",
"group": "What agents can do",
"year": "2026",
"lead_author": "Li",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-what-agents-can-do",
"contribution": "Supplies a reusable 932791-PR universe and a curated 33596-PR subset that underpins many agent outcome studies.",
"limitation": "Shared data are not independent replications; five agents and snapshot through mid-2025.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "https://arxiv.org/html/2602.09185",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"audit_locator": "arXiv 2602.09185 html Introduction and dataset description; chapter 7 earlier audit primary-source register",
"data_snapshot_warning": "Dataset observational and shared by W068/W071; overlapping observations are not independent replication.",
"previous_chapter_audit_registers": "CH03:ARXIV_ABSTRACT_VERIFIED; CH07:FULLTEXT_HTML_DATASET",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}W193PolimiIn the published NDSE2 bibliography
METR / Epoch AI · 2026
Agents can replicate substantial multi-file implementation work in a constrained exercise, challenging simplistic duration ceilings.
Evidence, context & provenance
Evidence boundaryHuman-hours counterfactual inferred; not general work-hour substitution or deployment readiness.
Source type, as recorded: Observational / qualitative evidence
Editorial use. Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.
Author names can be lead-author fields, not complete bylines. Follow the recorded source for the full citation. An index URL is labelled as an index, not silently replaced by an unverified primary-paper link.
Topics. Controlled productivity evidence · Evaluation methodology and baselines · Long-running agent execution
11 recorded source relationships
These retain the graph’s exact relation labels. A connection is not automatically a causal claim or an independent replication.
All preserved graph attributes
{
"label": "MirrorCode: Evidence that AI can already do some weeks-long coding tasks",
"kind": "paper",
"group": "What agents can do",
"year": "2026",
"lead_author": "METR / Epoch AI",
"source": "https://www.sepolimi.com/p/se-ai-bibliography-what-agents-can-do",
"contribution": "Agents can replicate substantial multi-file implementation work in a constrained exercise, challenging simplistic duration ceilings.",
"limitation": "Human-hours counterfactual inferred; not general work-hour substitution or deployment readiness.",
"manual_use": "Make the engineering response an explicit scoped permission, testable gate or human escalation criterion.",
"evidence_class": "Observational / qualitative evidence",
"editorial_priority": "B",
"bibliography_year": "2026",
"original_evidence_class": "Observational / qualitative evidence",
"reviewed_evidence_class": "Observational / qualitative evidence",
"full_manuscript_audited": "no",
"cross_domain_boundary": "",
"publication_year_from_independent_source": "",
"publication_source_url": "",
"publication_status_note": "",
"section_id": "",
"summary": "",
"method_status": "",
"provenance": "",
"node_class": "LEGACY_EDITORIAL_GRAPH",
"baseline_provenance": "2026-10-09 corrected editorial crosswalk",
"previous_chapter_audit_registers": "CH03:NOT_INDEPENDENTLY_INSPECTED; CH07:FULLTEXT_INSTITUTIONAL",
"previous_chapter_audit_count": "2",
"previous_audit_warning": "Archived self-reported inspection depth; no independent reinspection from these archive records in 2026-10-10 pass.",
"ndse2_actual_chapters_json": "[\"C:C03\", \"C:C07\"]",
"ndse2_in_bibliography": true,
"ndse2_citation_basis": "Final 13.7.1 active EN/ITA source tree; index links and nocite are not chapter citations"
}