# Adversarial review log - super-enhancer-cell-identity (shipped review)

Canonical file: `super-enhancer-cell-identity.review.md` (batch `hsef_v53_final_nine_2026-10-12`).
This log is named for the slug because the same folder already holds `ADVERSARIAL_REVIEW.current-source-checkpoint.md`,
the log of a separate pass over the build intermediate `current-source-checkpoint.review.md`.

Level before -> after: Β -> Β (unchanged). Confidence Moderate -> Moderate (unchanged; rationale rewritten under lead confidence-band policy).
Findings: 1 Major, 3 Minor, 0 Editorial. PMIDs verified 66/66. Claims spot-checked 10. Validator: PASS (0 fail, 0 warn, 31 pass).

**Carried-forward item resolved.** The earlier checkpoint pass flagged, without resolving, a tally mismatch: prose "10 screened
discriminating contrast units" against eight printed inventory items. Resolved here from the score card's `evidence_inventory`:
`decisive_perturbation: true` holds for exactly the eight printed records ([4], [5], [6], [8], [9], [10], [36], [61]) and
`discriminating_contrast: true` holds for ten (those eight plus [3] and [7]). Both numbers are correct and measure different
sets; no number was changed and a clarifying clause was added instead. The same resolution applies to the build intermediate
and to `causal_tally_audit.screened_discriminating_contrasts` in the score card, which needs no change.

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1. **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-10-14.

**Edition note.** This is the shipped review. The `current-source-checkpoint.review.md` file in the same folder is a build
intermediate carrying the same hypothesis in a later source-card edition; it was adversarially reviewed separately and is marked
superseded. The two files are not byte-identical and differ in the registered-claim wording, the selection provenance and the
sections they carry, so findings were re-derived against this file rather than copied across.

**Scope of checks.** All 66 REFERENCES PMIDs verified against PubMed `get_article_metadata`: 66/66 resolve to the stated first
author, journal, year and DOI; 0 failures; no retraction or expression-of-concern flag. Ten load-bearing claims were
spot-checked against the cited abstracts, concentrating on the native perturbations that hold the class claim at Β: the
alpha-globin five-element deletion series and its additive, non-synergistic result; the embryonic stem cell cluster deletions
with their 12 to 92 per cent range and the strong enhancers that super-enhancer calling misses; the Wap motif hierarchy with
additive tested pairs; the facilitator reconstruction showing interchangeable, position-dependent elements; the cluster
inversion with boundary and promoter controls; the engineered contact-loss experiment that preserves RNA; the Vsx2 retinal
deletion that removes bipolar neurons; and the BET inhibition and degradation result that separates transcription from retained
enhancer-promoter contact. Every checked statement matched its source. Scoring logic was re-derived: caps precede aggregation;
the declared `weakest-link` rule gives min(Β, Β, Γ) = Β, which matches the printed headline; the Β/Γ/Δ boundary and the Ζ
independence basis were checked.

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| S1 | Major | Decisive-perturbation inventory, the paragraph immediately after the tally | "Pivotal symptom studies remain informative; missing registration or mediator verification is an audit limitation, not evidence that dopamine has no effects or that antipsychotics lack efficacy." to a sentence about native-locus perturbations, activity-matched comparators and class-specific mediation | Template bleed from an unrelated dopamine case. A neurotransmitter and a drug class appear nowhere else in this review and have no bearing on its claim; the sentence asserted a boundary about antipsychotic efficacy inside a super-enhancer review. The same bleed was found and corrected in the sibling build intermediate. |
| S2 | Minor | Decisive-perturbation inventory, tally sentence | Clause added stating that the ten discriminating contrasts are a different set from the eight listed decisive perturbations, and naming records [3] and [7] as the difference | Reported as an unresolved count mismatch by the earlier pass on the build intermediate. Resolved here against the score card: `evidence_inventory` carries `decisive_perturbation: true` for exactly eight records ([4], [5], [6], [8], [9], [10], [36], [61]), which are the eight printed, and `discriminating_contrast: true` for exactly ten ([3] and [7] in addition). Both numbers are therefore correct and neither was changed; the defect was that the prose did not say they count different things. |
| S3 | Minor | Final score and M/S/D role derivation, revision-checkpoint blockquote, REVIEWER NOTES, and four Access/status lines | "(step1)" to "(step 1)" through "(step5)" to "(step 5)"; "three Step15 scoring blocks" to "three Step 1-5 scoring blocks"; "Prior ranks43/44 and rank49" to "Prior ranks 43/44 and rank 49"; "ErratumIn:" to "Erratum In:" | Known renderer defect that drops spaces around numerals. No content added or removed. |
| S4 | Minor | `final_score.confidence_rationale` (scorecard only; not printed in the prose) | Rationale leaning on "incomplete access" to a rationale built on evidence density and reproducibility, with access depth struck | Lead policy: access depth is not evidence density and never justifies a band. Re-deriving from what remains leaves an additive alpha-globin series, cluster deletion effects ranging from dispensable to near-essential, strong non-called enhancers, interchangeable rather than class-specific facilitators, unresolved phase specificity, and few activity-matched comparators. Those are real reproducibility and control-design reservations, so **Moderate stands**. |

**Rejected findings (checked, no change).**
(i) "Best-supported rival" label: `competing_models.models_evaluated[0]` is the rank-1 model with `is_nominal: false`, so the
label and the printed rank order are both correct. The renderer defect seen in sibling reviews is absent here.
(ii) Weakest-link aggregation: min(Β, Β, Γ) = Β is correct, and the derivation names SC1 and SC2 as the binding conjuncts.
(iii) Momentum "Mechanistic only": the artifact states that the recent reconstruction, inversion and imaging papers are completed
reports and that no named ongoing class-discriminating experiment is asserted. Correct value, and it does not rest on a
completed publication being treated as work underway.
(iv) No collective-author defect of either known form. Every first author in the reference list is a personal surname; there is no
leading-comma instance and no organisation fragment such as "Group", "Network" or "Consortium" standing in for a consortium
name.
(v) Reference 66's title is printed with the italicised "Eμ" that the PubMed title field drops. The review is more accurate than
the metadata here; not an error.
(vi) PubMed title fields dropping italicised gene or marker names at references 15 and 24. Declared non-error.
(vii) The banner's "54 complete abstract/status examinations" differs in wording from the sibling house phrase "bounded
abstract/status examinations". Checked: the document consistently claims complete *abstracts*, never complete papers, and the
REVIEWER NOTES record zero complete-paper or supplement audits. Not an overclaim; left as written.
(viii) No spliced "The earlier adversarial approval refers to its original edition" sentence is present.
(ix) Levels: SC1, SC2, SC3 and the headline Β follow from the quoted caps and the declared aggregation rule. No letter required a
change.

**Scorecard change.** `final_score.confidence_rationale` only. The level, confidence, subclaim levels, tally and every other
field are unchanged.

**Residual limitations.** Twelve of 66 records were examined at partial-body depth; the rest are abstract or status
examinations, as the banner states. Five requested targets were unavailable, as the REVIEWER NOTES record. The scope-matched
sensitivity is reported as Β for a conjunctive registered claim whose conjunction no single subclaim's scope covers; this was
argued both ways and left as published, and it does not affect the headline, which the preregistered weakest-link rule sets.

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



## Lead-adjudicated reconciliation (Editorial, 2026-10-14)

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| S5 | Editorial | `current-source-checkpoint.review.md`, banner and `## ADVERSARIAL REVIEW AND APPROVAL` heading, in both copies | Build intermediate carried a plain adversarial banner and an approval section with no indication that it is not a review, alongside the superseded note added in this pass | Lead adjudication. An earlier group reviewed this checkpoint in the belief that it was a review, before `current-source-checkpoint.review.md` was established to be a build-tool working filename. That pass was careful and its findings are real - the template bleed corrected here as S1 and the tally item resolved here as S2 both came from it. The error was in the inventory, not in that group's work. Deleting the approval section would erase the record of a review that genuinely happened, which is the same class of error as the false-approval boilerplate this round is correcting. The banner and section were therefore kept and amended: both now state at the top that the file is a superseded build intermediate, that the pass was performed on the intermediate before that was understood, and that it does not constitute approval of the Super-enhancer Model of Cell Identity review, naming `super-enhancer-cell-identity.review.md` as the review that is approved. Applied identically to both copies and hash-verified. |

The authoritative banner and approval for this case are on the shipped `super-enhancer-cell-identity.review.md`, from this pass.