# Adversarial review log — dohad-barker

Authoring model: **Sol 6.1**. Adversarial reviewer: **Claude Opus 5.5 (Anthropic)**. Date: **2026-10-14**.
Round 2 of the adversarial pass (brief: `deliverables/ADVERSARIAL_REVIEW_BRIEF_2026-10-06.md`, ROUND 2 section).

| Item | Value |
|---|---|
| Level before | Ε — Context-specific causality established |
| Level after | Ε — unchanged |
| Confidence | Moderate — unchanged |
| Role | reinforcing / stabilizing factor — unchanged |
| M / S / D | Mechanistic only / Stable / Contested — unchanged |
| PMIDs checked | 66 / 66 resolve to the stated title, first author, journal and year |
| PMIDs failed | 0 |
| Retractions / EoC | none found |
| Load-bearing claims spot-checked | 14 |
| Findings | 0 Critical, 0 Major, 0 Minor, 1 Editorial (16 instances) |
| Validator | PASS — 0 fail, 0 warn, 32 pass under the archived v5.3 validator (`validator.txt`) |

## Findings

| ID | Severity | Location | Change |
|---|---|---|---|
| B1 | Editorial | Throughout (16 instances) | Rendering defects: missing spaces around numerals ("at30", "The2026fasting", "at least2.5kg" ×3, "quoted5.16diabetes", "issue-year2026", "provider's2025dating", "online2025", "The2019birthweight", "updated2024selection", "step5", "atΕ"); bare PMIDs given their prefix ("follow-up25537745", "(11118027)", and the four Momentum citations); empty "DOI: ." at reference 50 → "DOI: not assigned in PubMed." (PubMed carries no DOI for PMID 10369801). |

No substantive error was found. This is the cleanest review in group 8: every spot-checked number and design detail matched its source, and the authoring model had correctly caught several traps (the "high birth weight" category in PMID 39913677 meaning only ≥ 2.5 kg; its 5.16 diabetes hazard ratio being the observational Cox estimate rather than the MR result; the Hyderabad trial being non-randomised and extending to age six; the Dutch famine coronary finding resting on 68 early-gestation subjects with mid- and late-gestation strata not elevated).

## Systematic artifacts reported by the lead — all four checked

1. **False "earlier adversarial approval" boilerplate** — **absent.** This is a fresh review, not a 9 October revision; it carries no spliced Consensus-consistency block and no "earlier approval" clause.
2. **Confidence set from access depth** — **present as wording, band confirmed correct.** The score card's `confidence_rationale` names reviewer access alongside evidence density, although the framework defines confidence by evidence density and reproducibility. The band was independently re-derived against the High text and **Moderate is correct**: the human causal evidence rests on a small number of natural experiments and within-family comparisons with several nulls, and the animal literature is replicated but mechanistically heterogeneous. **Superseded 2026-10-14:** the lead has since directed that the rationale be rewritten per case; that edit is finding B2 and is described under “Lead policy applied” above. The band is unchanged.
3. **"Best-supported rival" renderer bug** — **absent.** `models_evaluated[0]` is "Exposure-specific programming with postnatal modification" with `rank_in_tournament: 1` and `is_nominal: false`; the nominal ("Fetal nutritional and hormonal programming") is the second entry. The prose names the correct rank-1 rival.
4. **Reviewer-notes / Consensus-splice contradiction** — **absent.** No spliced block exists, and the REVIEWER NOTES section affirms Consensus use consistently with the reconciliation section.

Score card `review_metadata.reviewer` left as issued; corpus-wide house provenance.

## Lead policy applied — confidence bands (2026-10-14)

Two-step test as issued by the lead. **Step 1, strike access depth as a determinant.** The score card's `confidence_rationale` named reviewer access ("incomplete access set the confidence"). How much of a paper the reviewer could read is neither evidence density nor reproducibility and never justifies a band, so the sentence was rewritten to drop it and to say explicitly that access depth is recorded per ledger row and is not a determinant. **Step 2, re-derive from what remains.** What remains is substantive: controlled prenatal perturbations replicate across rat and sheep contexts with window, sex and pair-feeding controls, but the human causal evidence rests on a small number of natural experiments and within-family comparisons alongside independent nulls, with heterogeneous exposures, gestational windows and endpoints across the 66 reports. The review also documents a material rubric-fit limitation. That is not the High text's extensive cross-modality replication. The band therefore **stands at Moderate**; access depth was not the only thing holding it down.

**Momentum re-checked against the same clarification** (a completed publication or trial never satisfies "Decisive work underway"): already "Mechanistic only"; no change needed.

## Rejected findings (checked, no change)

1. **Ε rather than Ζ** — the convergence across controlled animal perturbation, timed famine shock and discordant-sibship comparison is real, but the classes do not narrow to a common exposure-specific mediator at the broad human scope, and the nulls bound generalisation. The review's own steelman names that blocker.
2. **Ε rather than Δ** — the window, sex, strain, pair-feeding and cross-fostering controls map the context rather than leaving generalisation untested, which is the Ε definition.
3. **Citation years** on references 7 and 28 differ by one from the PubMed e-pub year — rejected per convention.
4. **Reference 34's date range** ("2005 May 28-Jun 3") matches PubMed — not an error.
5. **Nominal hypothesis shown as tournament Rank 2** — the score card flags `is_nominal: true` and records that the tournament did not move the level; ranked presentation is house style.
6. **Divergence "Contested" without arithmetic** — Contested is defined as the case where no single implied level applies.
7. **Momentum "Mechanistic only"** — artifact named (PMID 42120553, 39358993, 39913677, 38839195).
8. **Collective-author reference defect** — none present.
9. **Absent lay summary** — mandatory only from framework v5.4; this is a declared v5.3 artifact and passes the archived v5.3 validator. Recorded as a residual limitation, not a correction.

## Files changed in this case folder

- `dohad-barker.review.md` (banner + ADVERSARIAL REVIEW AND APPROVAL section + B1 corrections)
- `validator.txt` (new)
- `SHA256.txt` (refreshed for the changed review file)
- `sol61-original/` holds the pre-review copies of the review, score card and SHA256.txt.

`dohad-barker.scorecard.json` was **not** changed — no scoring field needed correction.
Package-level files (`README.md`, `MANIFEST_SHA256.json`, `READER.html`, `COMPLETION.json`) were **not** edited.

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