# Adversarial review — metformin-geroprotection (ROUND 2)

Authoring model: **Sol 6.1**. Adversarial reviewer: **Claude Opus 5.5 (Anthropic)**. Date: **2026-10-14**.
Canonical file reviewed: `consensus_consistency_audit_2026-10-09/revised/metformin-geroprotection/metformin-geroprotection.review.md`
(the revised edition; reviewed as it now stands).

## Summary

| Item | Result |
|---|---|
| Level before → after | Γ → Γ (no change); SC2 remains Ε, reported at its own scope |
| Confidence before → after | Moderate → Moderate (no change) |
| Causal role | epiphenomenon (unchanged) |
| Findings | Critical 0 · Major 0 · Minor 3 (4 string occurrences) · Editorial 0 |
| REFERENCE PMIDs verified | 61/61 resolve; 0 failures; 0 retraction/EoC flags |
| Extra records introduced by the revision, verified | 4/4 (PMID 42023167, 33428709, 36281624, 42620010) |
| Load-bearing claims spot-checked | 12 |
| Collective-author builder bug | not present (reference #14 has no PubMed author list; the group name sits in the title field) |
| Validator | package-stored report (v2.3) ok=True; re-run with the bundled v5.4 copy gives 1 FAIL (`v5.lay_summary`, a v5.4-only check on a v5.3 review), 0 warn, 31 pass — `validator.txt` |
| Regression vs 2026-10-09 edition | one, Minor: see finding 1 |

## Findings

| ID | Severity | Location | Original → corrected |
|---|---|---|---|
| 1 | Minor (regression) | REVIEWER NOTES | "…Consensus… was performed." → the claim scoped to the original review plus a clause identifying the 9 October 2026 Consensus consistency validation as a later provenance supplement |
| 2 | Minor | Evidence cards [12] (PMID 19878986) and [14] (PMID 9742977) | "ErratumIn: PMIDnot supplied." → "Erratum in: PMID not supplied." (×2) |
| 3 | Minor | REFERENCES #14 (PMID 9742977) | "DOI: . Access: abstract_metadata." → "Access: abstract_metadata." |

Finding 1 is the only regression: the revision added a Consensus consistency validation section while the
preserved REVIEWER NOTES still asserted that no Consensus work had been done, so the document contradicted
itself. The pre-audit edition had no such contradiction because it had no Consensus block. The fix scopes
the original sentence rather than deleting it.

## Verification of the block the revision added

Every quantitative claim in the 2026-10-09 supplement was re-derived from the cited records and is correct:
METFORAGING's 19/21 randomisation of 40 participants, 35 completers at 96 weeks and its primary PhenoAge
estimate of −1.02 years (95% CI −5.30 to 3.26, p = 0.627); the animal-lifespan meta-analysis finding no
significant overall effect in mice or nematodes with a live-OP50 worm subgroup signal; and the medRxiv
frailty trial's 145 randomised participants, its year-1 Fried deterioration attributable to the weight-loss
item, and its frailty-index change of −0.006 ± 0.0026 per year (p = 0.0222). The block's framework handling
of the preprint is also correct: it refuses to let a non-peer-reviewed report cross a threshold and records
that a judgement resting on it would force confidence Low. Observation only: the block omits that the same
preprint reports reduced epigenetic age on two clocks — biomarker evidence the review discounts elsewhere,
so an omission rather than a misstatement.

## Rejected findings (checked, no change)

1. Aggregation: primary-subclaim with SC3 named; headline = SC3 Γ. Alternatives correct (weakest-link
   min(Γ, Ε, Γ) = Γ; scope-matched Γ, since SC3 is the only subclaim scoped to non-diabetic human aging).
2. Causal role "epiphenomenon" is correctly derived at step 2 and the derivation says so.
3. Momentum "Mechanistic only" names its artifact and correctly declines "Decisive work underway".
4. Causal tally: 21 claimed, 21 bracketed items printed, score card agrees.
5. Reference titles at #30 and #42 differ from PubMed only by PubMed dropping an italicised species name
   and an italicised "via".
6. Reference #14's absent author field reflects PubMed, which holds no author list for that record.
7. `v5.lay_summary` validator failure is a v5.4-only requirement on a v5.3 review.

## Copies

- `propagate_to`: none.
- **Divergent copy left untouched:** `hsef_v53_remaining_priority_01_05_2026-10-09/metformin-geroprotection/metformin-geroprotection.review.md`
  (the pre-audit edition). It carries findings 2 and 3 but **not** finding 1, which exists only because of
  the revision.

## Approval

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