Verification Summary

This document is the publication-grade audit trail of citation verification for the manuscript “VIP Cortical Interneurons: A Critical Reassessment of the Disinhibition Framework”. It accompanies the manuscript bundle as a transparency artifact and records the methodology, calibration audit, acceptance criteria, and final residuals of the citation-verification pipeline.

1. Scope¶

Verification covered every inline citation in the 13 body sections (01_introduction.md through 13_conclusion.md). The Methods section (Methods.md) carries no inline citations and is excluded from verification counts. The unit of verification is the citation triple:

{ section_filename, paragraph_idx, sentence_idx, claim_text, cite_key, paper_doi, paper_title, paper_authors, paper_year, role }

A claim sentence containing the multi-key form {cite:p}`A,B,C` is exploded into three triples (one per cite-key). Same-sentence repeated cite-keys are retained as separate triples by the EXHAUSTIVE_COUNT rule (Phase 15V) so that triple counts match cite-key occurrences in the source markdown. The total population is 3,816 triples spanning 414 unique cite-keys and 414 unique paper DOIs across 13 sections.

2. Pipeline (Phases 15-19)¶

3. Verdict Taxonomy¶

The Phase-16 verifier emits one of three verdicts per triple. Verdicts are conservative (in dubio pro reo: cite-key passes only if the verifier finds positive support).

• VERIFIED -- Bib metadata (title, authors, year, DOI, venue) provides sufficient evidence that the cited paper supports the claim role (supports / qualifies / contradicts / introduces) at the granularity asserted in claim_text.

• NEEDS_FIX -- Bib metadata is sound but the citation is misattributed, miscontextualised, or asserts more than the cited paper actually supports. Phase-17A remediation is required.

• INSUFFICIENT_EVIDENCE -- The verifier could not establish support from bib metadata alone within the verification budget. Outcome is treated as residual uncertainty, not as falsification.

The original gate_scope.json candidate taxonomy (VERIFIED / MINOR / CHIMERIC / HALLUCINATED / MISATTRIBUTED) was simplified during Phase 16 to this three-state set, which is what the LITREVIEW verifier skill operationally emitted.

4. Calibration Audit (Phase 16V)¶

Phase 16 was sharded across 4 LITREVIEW workers (W0-W3). The 16V validator computed per-worker INSUFFICIENT rates and flagged a methodological-stringency drift in W1 (54.6% INSUFFICIENT vs 9-18% in W0/W2/W3). This drift was not a parser bug -- W1’s verdicts were internally self-consistent, but its threshold for “sufficient evidence” was strictly more conservative than the others.

The 16V recommendation was to (a) prioritise the 435 NEEDS_FIX triples for body-text rewrite/recitation in Phase 17A, and (b) re-verify W1’s 521 INSUFFICIENT triples (batches B0053-B0105) plus the other workers’ 360 IE triples under W0/W2/W3 calibration with a body-text fetch step.

Phase 17B implemented this re-verification across all 881 IE triples: 351 flipped to VERIFIED, 127 flipped to NEEDS_FIX, and 403 remained INSUFFICIENT. The flip distribution confirms the drift hypothesis (351/881 = 39.8% under-credited verifications recovered, 127/881 = 14.4% genuine misattributions surfaced).

Phase 17B-V (aggregate validation) confirmed all 12 structural checks PASS: merged total 3,816, sum equals 3,816, all 13 sections present, NEEDS_FIX master count 562, no duplicate triple_ids, all triples have a final verdict, and per-state totals (V=2,851 / NF=562 / IE=403) match the sum.

5. Final Verdict Distribution¶

                         Phase 16    Phase 17B (final)
VERIFIED              :     2,500    ->     2,851
NEEDS_FIX             :       435    ->       562
INSUFFICIENT_EVIDENCE :       881    ->       403
                       -------       -------
total                 :     3,816          3,816

Final per-section verdict distribution (from verification_report_v2.json):

6. Top Problem Cite-Keys (after Phase 17B)¶

These were the cite-keys most over-cited or miscontextualised in the v5 manuscript and the primary targets of Phase-17A remediation.

7. Phase-17A Fix Application¶

The 562 NEEDS_FIX triples were sharded across 4 LITREVIEW workers (parallel fix-application):

Worker constraints (verified by Phase 17V):

• Zero new cite-keys introduced (all post-fix keys remain a subset of the 414 in the references.bib allowlist).

• Word-count drift within +/-5% per section (max observed -1.18%).

• No new forbidden lexicon introduced (the 35 residual canonical/scaffold hits in 17V are legitimate scientific usages such as “canonical disinhibitory motif” and “developmental scaffold”; 0 process-language hits).

• myst build continued to succeed cleanly (14 pages, 0 errors, 0 warnings).

8. Phase 18 Book-End Critic and Remediation¶

Phase 18 (cross-section consistency critic, LITREVIEW) issued FAIL on iter1:

• 1 MUST_FIX (MF-18.1): an orphan-quoted RNAscope/IHC statistic in sec_11 (“VIP+ interneurons corresponded to 4.5, 5.2, and 6.6% ...”) had lost its attributing cite-key in a Phase-17A REMOVE_CITE pass; the Teymornejad2024a citation was restored.

• 5 SHOULD_FIX items covering stylistic repetition (Prior reports x61 globally, An earlier study x35 globally, in the cited x7 in sec_07), one meta-jargon leak (citation key map x6 in sec_11 -- removed and rephrased as allowlisted reference set / allowlisted bibliography), and one antecedent-disambiguation in sec_05 (replaced “The same paper organised hippocampal IS interneurons ...” with “Guet-McCreight et al. (2020) further organised ...”).

The Phase-18 remediation send-back (DATAML) executed 111 edit operations across 8 sections, addressed all 1 MUST_FIX + 5 SHOULD_FIX items, and verified zero new cite-keys, +/-5% word-count drift, and Phase-17A action preservation.

9. Phase 19 Final QA¶

Phase 19 ran a 16/16-check structural QA on the v6 manuscript (post Phase-18 remediation patch on the v5 baseline). All 16 checks PASS:

10. Acceptance Criteria¶

Acceptance for the manuscript bundle is the conjunction of:

• Phase 14V: 16/16 structural assembly checks PASS.

• Phase 15V: 9/9 triple-extraction structural checks PASS.

• Phase 16V: 9/9 verification sanity checks PASS; calibration audit produces an actionable recommendation that Phase 17B implements.

• Phase 17B-V: 12/12 verification-aggregate checks PASS.

• Phase 17V: at minimum 5/6 structural checks PASS; FORBIDDEN_LEXICON 17V conditional pass requires every residual hit be a legitimate scientific usage and a 0-count on process-language tokens (orchestrator, verdict). Both conditions hold (35 lexicon hits all legitimate; 0 process-language hits).

• Phase 18 + 18-rem: 1 MUST_FIX + 5 SHOULD_FIX from the consistency critic resolved; constraints satisfied (zero new cite-keys, +/-5% word-count drift, no new forbidden lexicon, Phase-17A actions preserved).

• Phase 19: 16/16 final QA checks PASS.

• Phase 20 (this phase): Methods ledger refreshed; verification summary published; manuscript pushed to GitHub; final delivery tarball assembled.

All criteria are met. The Phase-20 final gate (gates/gate_phase20_final.json) records PASS with the qualifier that the bibliography target of >=2,000 papers from the original review request was not met (1,038 entries, 391 actively cited) -- this is documented as an accepted Phase-2V outcome of the saturation rule (which exited at 1,037 unique DOIs).

11. Residual Uncertainty: 403 INSUFFICIENT_EVIDENCE Triples¶

The 403 triples that remained INSUFFICIENT after Phase 17B re-verification represent claim/citation pairs for which body-text evidence could not be retrieved within the verification budget. They are not falsified citations: the verifier’s threshold is “positive support found” (in dubio pro reo). These triples concentrate in the dense circuit-mechanism sections (sec_06 synaptic properties: 104; sec_07 disinhibition: 79; sec_12 computational models: 58; sec_05 electrophysiology: 38), reflecting the high citation density of those sections rather than a manuscript-quality issue.

These residuals are catalogued in verification_report_v2.json (per-triple verdict_final="INSUFFICIENT_EVIDENCE"); a future review iteration can target them with a deeper body-text fetch and reverification pass.

12. Reproducibility¶

The verification artifacts are reproducible from the inputs and skill protocols listed in the table at Section 2. Specifically:

• Phase 15 triple extraction: comprev-triples-extractor skill applied to manuscript v5 content/*.md and references.bib (1,038 entries).

• Phase 16 verification: 4 parallel LITREVIEW workers running comprev-triples-verifier over 212 batches of 18 triples.

• Phase 17B re-verification: 3 LITREVIEW workers re-running the verifier with body-text fetch on the 881 IE triples.

• Phase 17A fix application: 4 parallel LITREVIEW workers running comprev-fix-application over per-worker NEEDS_FIX shards.

• Phase 18 critic: single LITREVIEW frame running comprev-critic in book-end mode.

• Phase 18 remediation, Phase 19 QA, Phase 20 assembly: DATAML mechanical work.

The full per-triple verdict file is verification_report_v2.json (5.6 MB; 62,949 lines; one record per triple). All gate JSON files are in gates/.