# Adversarial review - ultra-processed-food

**Authoring model:** Sol 6.1 (original edition), with a cross-model red-team synthesis already recorded in the document. **Adversarial reviewer for this pass:** Claude Opus 5.5 (Anthropic). **Date:** 2026-10-14.
**Level before:** Gamma (Mechanistically plausible, partial system support), Moderate confidence.
**Level after:** Gamma (Mechanistically plausible, partial system support), Moderate confidence. Unchanged.
**Momentum before:** Decisive work underway. **Momentum after:** Mechanistic only (lead-adjudicated).

## Summary

| Severity | Count |
|---|---|
| Critical | 0 |
| Major | 0 |
| Minor | 2 |
| Editorial | 0 |

PMIDs verified: 28 reference PMIDs plus the 2 attached errata, 30/30 resolving to the stated record, 0 failures, no retraction or expression-of-concern flag. Item [18] is a ClinicalTrials.gov results record and is described as such.
Load-bearing claims spot-checked: 13, all matching the cited abstracts exactly, including every quantity the verdict turns on.
Validator after editing: PASS under the registry-compatibility build of validator 2.4 that this package documents and validates under (0 fail, 1 warn, 33 pass, matching the package's own validation-report.json). The base 2.4 build fails only on the official registry results row for lacking a PMID or DOI, and fails identically on the pre-edit original.

This is the cleanest review in group 9 on citation and rendering integrity: no rendering defect, no empty DOI field, no aggregation misread, and a Divergence value that is actually computed, with both implied levels and their arithmetic. Its one substantive correction is the Momentum value, adjudicated by the lead.

## Findings

| ID | Severity | Location | Correction |
|---|---|---|---|
| E1 | Minor | Consensus provenance reconciliation block; consensus-addendum.md | "The earlier adversarial approval refers to its original edition" rescoped: the 2026-10-06 approval attached to this case covers the dated 2026-10-05 research addendum, not this review edition, which carried no earlier approval |
| E2 | Minor (lead-adjudicated) | AT A GLANCE; section 13 Momentum; score card momentum block, change record and red-team snapshot | Momentum "Decisive work underway" corrected to "Mechanistic only". The framework requires that someone be running an experiment that could cross the next threshold; NCT05290064 has completed, so nothing is being run, and its posted contrast is non-significant and would not carry SC4 across Delta. Naming a real trial meets the artifact-specificity requirement but not the definition. Momentum modifies no letter. |

## Rejected findings

Confidence Moderate, whose rationale is density-based and does not use access depth; the "generated CRISPR caption" clause, which looks like imported wording but is verified true against this case's own 2026-10-05 enrichment records; the best-supported-rival renderer defect (no such line in this edition, and models_evaluated[0] is a non-nominal rank-1 model); weakest-link aggregation with both alternatives reported and scope-matched correctly scored on SC4 and SC5; the reviewer-notes contradiction (not present); the base-validator registry-row FAIL (pre-existing, and the reason the documented compatibility build exists); the v5.study_identity warning (pre-existing and recorded); the structural placement of the spliced block.

## Approval

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