# Adversarial review log — shingles-vaccination-dementia

Authoring model: **Sol 6.1**. Adversarial reviewer: **Claude Opus 5.5 (Anthropic)**. Date: **2026-10-14**.
Round 2; group 7 of `NEW_REVIEW_INVENTORY_2026-10-14.json`.

## Summary

| Item | Value |
|---|---|
| Level before | Δ — First field-valid causal perturbation crossed / Moderate |
| Level after | Δ — First field-valid causal perturbation crossed / Moderate (unchanged) |
| Findings: Critical | 0 |
| Findings: Major | 2 |
| Findings: Minor | 2 |
| Findings: Editorial | 1 |
| Rejected findings | 7 |
| PMIDs checked / failed | 53 / 0 |
| Load-bearing claims spot-checked | 12 (every reported effect size) |
| Validator | FAIL (1 fail, 0 warn, 33 pass) — `v5.lay_summary` only; identical before and after. |
| Approval | **Approved as revised.** |

## Findings

| ID | Severity | Location | Original → corrected | Basis |
|---|---|---|---|---|
| AR-1 | Major | Prose review (100 tokens) and score card (228 tokens) | Numeral-adjacent space loss repaired, including grade-bearing statistics: "effect3.5percentage" → "effect 3.5 percentage", "diagnosis1.8points" → "diagnosis 1.8 points", "with95%CI0.0–54.7" → "with 95% CI 0.0–54.7", "despite57covariates" → "despite 57 covariates", "not17%risk" → "not 17% risk", "age80" → "age 80". Dropped leading zeros restored to match the sources ("HRs.89/.68" → "HRs 0.89/0.68", "CI−.40to.27" → "CI −0.40 to 0.27", "RD−.06" → "RD −0.06"). 38 bare reference numbers fused to the preceding word given their "PMID" prefix, matching the review's own convention | Known "randomize3808" defect class, here degrading statistics rather than only prose. Identifiers (PMID, PMC, NCT, DOI, study_id, receipt handles, version strings, hashes, URLs, AS01/HSV/HbA1c) guarded and verified unchanged; whitespace-insensitive comparison confirms nothing else changed. |
| AR-2 | Minor | Final-score block Divergence line; score card divergence justification | Artifact added naming the two conflicting framings | Divergence requires a checkable artifact; both prose and card asserted only its absence. Value unchanged at Contested. |
| AR-3 | Editorial | Same fields | Original disclaimer retained alongside the new artifact | No loss of the author's qualification. |
| AR-4 | Major | AT A GLANCE, "Best-supported rival" line | Nominal ("Local vaccine effect on dementia diagnosis", rank 1) → "Health selection and ascertainment (rank 2)" | Renderer defect: the "Best-supported rival" line is generated from `competing_models.models_evaluated[0]`, which here carries `is_nominal: true` and rank 1, so the prose named the nominal as its own rival. Prose-only; the tournament, the level and Rule 9 are untouched. The paragraph body already argued the Welsh design is the strongest separator against healthy-user selection, i.e. against the rank-2 rival, so the corrected label makes it read as intended. |
| AR-5 | Minor | Score card `final_score.confidence_rationale` | Access depth removed as a confidence determinant | The framework ties confidence to evidence density and reproducibility; the boilerplate rationale named "incomplete access" as a determinant. Rewritten to rest on density, replication and heterogeneity. Band re-tested against both band texts and unchanged. Moderate stands: the causal core is a small family of eligibility natural experiments with an unresolved published challenge. |

## Rejected findings (checked, no change)

1. **Reduction to Β.** Argued on the 2026 quantitative critique (implied effectiveness of 177%, 84% and 111% outside Wales; no population-level dementia decline) and the unreviewed English null. Rejected: the framework names natural experiments as a field-valid equivalent, and Δ requires one such perturbation with a positive, phenotype-relevant result. The Welsh first stage runs 0.01% → 47.2% across one week of birth date with balanced covariates and placebo cutoffs; the critique attacks magnitude and cross-country coherence rather than that local design, and the English analysis is a preprint the review already excludes from its count. Ε also rejected — the sex boundary rests on a non-significant interaction, so no causal failure context is mapped. Most contestable judgement in the review.
2. Aggregation — primary-subclaim pre-registered, SC2 named, alternatives reported (weakest-link Β, scope-matched Δ).
3. Collective author — reference 50 renders correctly as "SPRINT MIND Investigators for the SPRINT Research Group, Williamson JD, …"; the recurring builder defect did not fire.
4. Citation years for refs [4], [12], [33], [49] — print/e-pub convention.
5. Momentum "Mechanistic only" with a trial-type artifact — thin but compliant, non-modal; no change.
6. Empty `DOI: .` fields, retrieval handles, "PMID not supplied" strings — none present.
7. Lead-reported artifacts screened: no false-approval boilerplate (not a revised edition), no reviewer-notes/provenance contradiction. The band was re-tested against both framework texts and Moderate stands.

## Numeric audit

Every figure matched its source, and the estimand discipline is this review's strongest feature: the Welsh receipt effect is reported as 3.5 percentage points over seven years rather than as the source's 20.0% relative reduction (40175543); the Australian figures are an eligibility effect of −1.8 percentage points on uptake of +16.4 points over 7.4 years, explicitly not a receipt effect (40267506); the Canadian clinical effect is about 2 percentage points over 5.5 years with uptake from a separate survey (41579903); the recombinant transition is "17% more diagnosis-free time and 164 additional days in those subsequently affected", stated twice as *not* a 17% risk reduction (39053634) — the most commonly misquoted figure in this literature, correctly handled; Optum HRs 0.89 and 0.68 (39733478); 509,926 entrants, 1.73% vaccinated, 87.0% after discharge, 57 covariates, negative controls showing residual confounding (42296498); AS01 18-month window with no between-product difference, and the source's own adjuvant-necessity inference refused (40562756); 326 villages, 33,995 participants, dementia risk ratio 0.85 over 48 months (40258956); and the critique's implied effectiveness figures reported as stated (42754179).

## Residual limitations

The Welsh analysis frame of 282,541, the Canadian first-stage estimates of 27.4 and 13.9 percentage points, and the sex-interaction p=.218 are Methods- or Results-level figures, confirmable in this pass only against the review's own partial-body receipts. Access depth unchanged (3 partial-body, 50 bounded abstract/status). Validator FAIL is a framework-version mismatch.

## Strings for the lead to propagate to package-level files

No level, confidence, role, source count or claim count changed. One scalar string does need propagation: the best-supported-rival label, old `Best-supported rival: Local vaccine effect on dementia diagnosis`, new `Best-supported rival: Health selection and ascertainment (rank 2)`. If OVERVIEW.md / READER.html / EDITORIAL_HANDOFF.md reproduce this review's prose or its numbers, re-render them from the corrected review rather than patching individual tokens — the AR-1 repair touched 100 tokens in the review and 228 in the score card, including figures a reader would quote.

**Lead policy recorded (confidence bands, 2026-10-14).** The lead adjudicated the question raised in this pass and set a two-step test for the corpus. Step 1: strike access depth as a determinant everywhere — how much of a paper the reviewer could read is neither evidence density nor reproducibility, and it never justifies a band. Step 2: re-derive the band from what remains; if access depth was the only thing holding the band down and the cited evidence meets the High text, raise it, but if the review documents substantive evidence-quality or reproducibility concerns (indirect effects, incomplete washout, residual confounding of the mechanism, reporter rather than endogenous readouts) those are valid determinants and the band stands. A reviewer's uncertainty about the evidence is a legitimate reason for Moderate; a reviewer's uncertainty about their own reading access is not. This pass's rationale rewrite is step 1; the band decision below is step 2. Step 1 applied (rationale rewritten under AR-5). Step 2: band stands at Moderate on valid determinants — the causal core is a small family of eligibility natural experiments with an unresolved published quantitative challenge, which is not extensive, well-established replication.

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