# Adversarial review - histone-code

Level before -> after: Γ Moderate -> Γ Moderate (unchanged). Momentum: Decisive work underway -> Mechanistic only.
Findings: 3 Major, 1 Minor. Validator: PASS (0 fail, 1 warn, 31 pass); the single warning is the standing-reject v5.claim_form 'phenomenon' notice.

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1. **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-10-14. This is the revised (9 October 2026) edition; the original edition was not adversarially reviewed, so this is the first adversarial review of this case.

**Scope of checks:** All 62 reference PMIDs re-fetched with the PubMed MCP tool get_article_metadata (title, first author, journal, year, DOI): 62/62 resolve to the stated record; no retraction or expression of concern. 11 load-bearing claims spot-checked against PubMed abstracts and, for PMID 38724747 (Policarpi), against the local full text in the originating batch: refs 5 (NURF, PMID 42165139), 6 (BPTF, PMID 21596426), 33 (KAT6B, PMID 39731917), 37 (TRIM33, PMID 22196728), 43 (H4 factorial, PMID 15795371), 49/50 (Set2/Rpd3S), 58 (Policarpi), 59 (PMID 36859550), plus the SC1 association sources. Scoring logic checked against the archived HSEF v5.3 framework, with particular attention to the declared scope-matched aggregation rule, the Zeta independence basis and the momentum value. Known upstream defects checked: collective-author dropping (none present), empty DOI fields (none), erratum "PMIDnot supplied" strings (none), rendering defects (six found, corrected).

**Findings and corrections**

| ID | Severity | Location | Original -> corrected | Evidence |
|---|---|---|---|---|
| HC-1 | Major | Consensus provenance reconciliation block (also consensus-addendum.md) | "The earlier adversarial approval refers to its original edition; ..." -> "...The original edition of this review carried no adversarial approval; the adversarial review of record is the one recorded in ADVERSARIAL REVIEW AND APPROVAL below." | The original edition, hsef_v53_popularity_ranks_36_40_2026-10-08/histone-code/histone-code.review.md, contains no adversarial banner and no ADVERSARIAL REVIEW AND APPROVAL section, and this slug is absent from the 2026-10-06 approved set. The 9 October revision asserted an approval that does not exist. This is a regression introduced by the revision, not an inherited error. |
| HC-2 | Major | AT A GLANCE table and the M/S/D paragraph (also scorecard extended_scoring_MSD.momentum) | Momentum "Decisive work underway" -> "Mechanistic only", with the artifact rewritten to say the named reports are completed publications and that no ongoing threshold-crossing work was verified | The artifact named only published outputs - Policarpi 2024, the 2025 bivalent/KAT6B report and the 2026 NURF report - and its own closing clause states that "no undisclosed ongoing trial or pending result is asserted". Under the lead's momentum policy, "Decisive work underway" requires work actually being run that could cross the next threshold, and a completed publication never satisfies it. Momentum modifies no letter, so no level or confidence change follows. |
| HC-3 | Major | scorecard final_score.confidence_rationale | "...9 selected partial-body examinations; density, heterogeneous controls and incomplete access set the confidence..." -> rationale resting on density and reproducibility: the SC3 combination evidence is reporter-based or recruitment-only, two of the four independence lines rest on a single paper each, and contradicting results (H3K27ac loss, cumulative H4 effects, the failed TAF3 switch) are unreconciled | Lead policy: reading access is not evidence density or reproducibility and cannot set a band. Re-derived under the two-step test, the band stays **Moderate** - the review documents substantive reproducibility and evidence-quality reservations, well short of the High text's "extensive independent replication across labs, models, and assay modalities". No level or confidence change. |
| HC-4 | Minor | Research-interpretation prose, M/S/D paragraph, evidence card [5]; mirrored in the scorecard | "Zhang2020", "Harris2023", "Policarpi2024" (x1 review, x2 scorecard), "Wang2023", "Local2018", "steps3/4", "readerdomain" -> spaces and hyphen restored | Known missing-space rendering defect. |

**Rejected findings (checked, no change)**

- Scope-matched aggregation: applied correctly. SC3 is the only subclaim whose scope matches the registered claim (combinations causally encoding outputs beyond additive individual-mark effects at endogenous loci) and it sets the Γ headline; SC2 at Ζ is narrower - individual marks, not combinations - and is reported separately, exactly as the rule requires. The published sensitivities (weakest-link Β, primary-subclaim Γ, scope-matched Γ) are consistent with the matrix. Note specifically that the Ζ subclaim is **not** allowed to set the headline here, which is the correct reading.
- "Best-supported rival": correctly rendered. models_evaluated[0] is M2, is_nominal false, rank_in_tournament 1; the nominal model (M1, "Strict combinatorial instruction") is correctly at rank 3 and is not presented as the rival.
- Attribution of PMID 38724747 as an "endogenous MLL2 catalytic knock-in with targeted mark restoration" was challenged and confirmed: the Policarpi body reports ESCs "with homozygous knock-in Y2602A catalytic point mutations (CM) in the H3K4 methylase gene Mll2 (Kmt2b)", 3,102 promoter H3K4me3 peaks lost and 15,244 retained. The review's parallel description of the same paper as a dCas9-based precision epigenome-editing platform with catalytic-dead controls is also correct; both arms are in the one paper.
- "NURF writer deletions/truncations" (SC3 and the cap paragraph) reads compressed but is accurate against the source: the deleted writers are Gcn5 and Jil1, and the truncations are of the BPTF/NURF301 reader domains, as the review states explicitly in the interpretation section. No change beyond the "readerdomain" spacing fix.
- KAT6B (PMID 39731917): "combination-specific recruitment plus KAT6B developmental loss" matches the abstract, which reports that the bivalent combination recruits proteins not recruited by either mark alone, including KAT6B, and that KAT6B knockout blocks neuronal differentiation. The review correctly declines to treat this as factorial mark necessity.
- H4 factorial (PMID 15795371): the review's "cumulative K5/K8/K12 effects versus special K16 behavior" matches the abstract's two-mechanism conclusion.
- TRIM33 (PMID 22196728) and BPTF (PMID 21596426): dual-mark recognition, PHD/Bromo architecture, H3K9me3 plus H3K18ac, and mononucleosome-level H4K16ac/H3K4me3 selectivity all match the abstracts.
- PMID 36859550: acute COMPASS-subunit ablation, KDM5-dependent turnover, pause-release and elongation rather than initiation - matches the abstract.
- The compact in-text PMID lists ("PMID38724747,36859550,16286007") are house style used consistently across this batch and were left unchanged.
- Print/e-pub year differences of +/-1 at refs 16, 33, 52, 54 and 60: acceptable under the stated convention.
- Validator warning `v5.claim_form` ("entry_form is 'phenomenon'"): rejected per the standing convention. The restatement grades a causal claim - combinations read as a code that causally specifies transcriptional outputs beyond individual marks - not a bare topic.

**Residual limitations**

- 53 of 62 sources were verified at abstract/status depth. Per lead policy this bounds what was checked and is not a determinant of the confidence band.
- Two linked Wang 2023 correction notices and the Local 2018 erratum remain unexamined, as the review already states.
- scorecard provenance_header.generated_by is unchanged; authorship is recorded here as Sol 6.1 per the mandate.
- Package-level files were not edited; propagation strings are listed in the agent report.

**Approval:** Following adversarial review and the corrections listed above, the revised document is approved. — Claude Opus 5.5 (Anthropic), 2026-10-14.