# Adversarial review log - laetrile-cancer-treatment

Authoring model: **Sol 6.1**. Adversarial reviewer: **Claude Opus 5.5 (Anthropic)**. Date: **2026-10-14**.
Round 2 of the Sol 6.1 HSEF v5.3 adversarial pass.

| Item | Value |
|---|---|
| HSEF level before -> after | A (Speculative / narrative) -> A (unchanged) |
| Confidence before -> after | Moderate -> Moderate (unchanged) |
| Causal role | epiphenomenon (unchanged) |
| Findings | Critical 0 - Major 0 - Minor 3 - Editorial 0 |
| PMIDs verified | 67/67 reference PMIDs plus 3 in-text identifiers; 0 failures |
| Claims spot-checked | 14 load-bearing claims against the original records; all matched |
| Validator | FAIL on `v5.lay_summary` only - the shipped validator enforces the v5.4 lay-summary rule against a self-declared v5.3 review. Cross-version artifact. 33 checks pass, 0 warnings. |

## Findings

| ID | Severity | Location | Correction |
|---|---|---|---|
| AR-1 | Minor | Throughout | ~96 renderer-dropped spaces repaired; bare numeric citations in AT A GLANCE, the competing-models block, M/S/D and REVIEWER NOTES given their `PMID` label for consistency with the rest of the document. Whitespace and label only. |
| AR-2 | Minor | REFERENCES 2, 7, 9, 20, 21, 25, 26, 27, 30, 32, 38, 39 | Twelve empty `DOI: .` fields replaced with "No DOI in PubMed record." (verified absent in the PubMed XML for all twelve). |
| AR-3 | Minor | scorecard `final_score.confidence_rationale` | Reviewer access depth removed as a stated input to the confidence band. Band unchanged (Moderate), and re-derived as correct against the framework's Moderate text. |

No level, confidence, role, momentum, stability, divergence, subclaim level, source count or claim count changed.

## Rejected findings

Headline A argued up to B on the six panel-judged 1978 responses (rejected: solicited-success sampling, 26 of 93 records unusable, no counterfactual, and cap 2 would bind in any case); the softness of the headline wording versus Moertel's own "not effective as a cancer treatment" (rejected - the refutation is carried by the grade, the "Additional downgrade: True" on the safety subclaim and the separate Zeta harm model, and the review correctly declines the one available overclaim, that amygdalin is inactive in every system); Zeta on the rank-1 harm model (independence basis recorded with four distinct failure modes and survives the v5.2 test); "Best-supported rival" (scorecard confirms rank 1, is_nominal false, nominal is rank 4 - renderer defect does not fire); Momentum "Mechanistic only" with four named artifacts; Divergence "Contested" (reviewer judgement, explicitly disclaimed as not a balanced-evidence claim); the "175 evaluable / one temporary response" figure (already correctly labelled secondary-source and listed as a knowledge gap); collective authors; rubric-02 boilerplate declared rather than silently resolved.

## Harm-and-refutation check (requested for this case)

Harm attribution is specific and correctly sourced. The cyanide evidence is not left as a generic warning: oral-versus-intravenous exposure is anchored to Moertel 1981 (oral dosing raising blood cyanide to 2.1 microgram/mL; intravenous drug largely excreted unchanged) and to Ames 1981; the mechanism of oral activation is anchored to Newmark 1981 (intestinal mucosal beta-glucosidase present, no detectable activity in the human neoplastic tissues examined) and Newton 1981 (gut-flora dependence, suppressed by antibiotics); causal specificity is anchored to Willhite 1982, where oral but not intravenous laetrile was teratogenic in hamsters and thiosulfate protected the embryos; and the clinical poisoning series is anchored to named case reports with measured cyanide, neurological injury, antidote response and the vitamin-C interaction. The review correctly separates massive apricot-kernel overdoses from the intended regimen, and correctly refuses to convert case reports into an incidence estimate. On the other side, it does not let the harm finding stand in for a benefit finding - the toxicity grade is kept out of the benefit grade, which is the right discipline and is stated explicitly. The one place a reader could be misled is the care-substitution subclaim, and the review already says in three places that those cohorts do not identify laetrile and that their hazard ratios are not drug-specific causal estimates.

## Files changed

- `laetrile-cancer-treatment.review.md` (banner, ADVERSARIAL REVIEW AND APPROVAL section, AR-1, AR-2)
- `laetrile-cancer-treatment.scorecard.json` (AR-3)
- `validator.txt` (new)
- `sol61-original/` holds the pre-review copies.

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