# Adversarial review log — h-pylori-peptic-ulcer

**Case:** h-pylori-peptic-ulcer (revised edition, `consensus_consistency_audit_2026-10-09/revised/`)
**Authoring model of record:** Sol 6.1. The document records editions A and B as graded by OpenAI Codex and this
edition as a Claude red-team synthesis of 2026-09-27; recorded, not edited.
**Adversarial reviewer:** Claude Opus 5.5 (Anthropic)
**Date:** 2026-10-14 · **Round:** 2, group 1

## Summary

| Item | Before | After |
|---|---|---|
| Headline level | Θ — High predictive accuracy, minimal unresolved exceptions | Θ — unchanged |
| Confidence | High | High — unchanged |
| Role | primary driver | unchanged |
| Subclaim levels | SC1 Θ · SC2 Θ · SC3 Η | unchanged |
| Aggregation | primary-subclaim (SC1); weakest-link Η, scope-matched Θ | unchanged |
| Momentum / Stability / Divergence | Little new work / Stable / Aligned (gap 0) | unchanged |
| Sources / studies / claims | 23 records · 22 studies · 34 claims | unchanged |
| Validator (5.4) | PASS — 0 fail, 2 warn, 34 pass | PASS — 0 fail, 2 warn, 34 pass |

**Findings by severity:** Critical 0 · Major 0 · Minor 1 · Editorial 1. Rejected findings: 8.

**PMIDs:** 23/23 verified against PubMed; 0 failures; no retraction or expression of concern in any record, which
independently confirms the review's own retraction screen.

**Claims spot-checked:** 12. Ten matched exactly; two produced the findings below.

## Findings

| ID | Severity | Location | Original → corrected | Evidence |
|---|---|---|---|---|
| C-1 | Editorial | "Limits and genuine alternative causes", and the matching claim-register row | "confirmed infected under a two-test definition" → "confirmed infected under a confirmatory definition — positive culture, or positive histology and CLOtest together, with at least two negative tests required to call a patient uninfected" | Ciociola et al. define infection as culture growth **or** both histology and CLOtest positive, and non-infection as at least two negative tests. "A two-test definition" implies two positives were always required. The 73%-of-2394 figure the sentence carries is correct and unchanged. |
| C-2 | Minor | "Clinical evidence and what it establishes" → "The finding generalizes beyond duodenal recurrence" | Added to the Graham sentence: "its triple therapy also contained bismuth subsalicylate, so the same mucosal-action caveat noted below applies, and the bismuth-free attribution rests on the duodenal trials." | The review flags the mucosal action of a bismuth-containing regimen for Sung in the next clause but not for Graham, whose triple therapy was tetracycline, metronidazole and bismuth subsalicylate. Those two are the review's only gastric-site sources, while the "narrow causal attribution" limb of its Θ case rests on the bismuth-free duodenal/NSAID-context trials — so the gastric generalization and the clean-attribution argument do not overlap. The review already concedes the parallel gap on the pathway limb; this makes the attribution limb consistent with it. |

Plus the mandated banner under the `> Sources:` blockquote and a new `## ADVERSARIAL REVIEW AND APPROVAL` section
before `## REFERENCES`.

**No level changed.** C-2 does not move a letter: SC1 is a disease-class claim the duodenal evidence alone supports,
and SC3 — the generalization subclaim — already holds at Η.

## The Η → Θ upgrade: argued against, then accepted

This was the central thing to test, and it survives. Θ requires "minimal", not zero, unresolved exceptions, and each
candidate second exception fails to qualify: the ~quarter of strictly tested US non-NSAID duodenal ulcers that were
uninfected bears on near-universality, which the graded claim explicitly disclaims; the Vergara null in prior NSAID
users is a context boundary determinable from baseline medication exposure, which the v5.3 scope rule permits; and
Cochrane's very-low quality rating concerns precision, not direction. HEAT's later window therefore remains the single
unresolved within-scope exception, it is carried by name rather than absorbed by redefinition, the penalty is taken
once (by SC3, not twice), and edition B's case for Η is retained verbatim as dissent. The one point where the review
could have been accused of circularity — the deleted "in susceptible hosts" qualifier, which is the exact case the
framework's v5.3 scope rule was written from — is handled correctly: deleted from the proposition and logged as
not determinable before the outcome.

## Rejected findings

R-1 … R-8 are tabulated inside the review. Besides R-2 (the upgrade, above), the ones worth the lead's attention: the
Θ headline drawn from a subclaim set containing an Η is **not** the aggregation misread the brief warns about (the rule
is primary-subclaim, pre-registered in edition A, and SC1 is coextensive with the graded claim; weakest-link's Η is
disclosed in three places); and the known upstream builder defects — collective author with a leading comma, fused
numerals, empty DOI fields, "PMIDnot supplied" — are all absent from this review.

## Files changed in this case folder

- `h-pylori-peptic-ulcer.review.md` — edited (2 corrections, banner, approval section)
- `validator.txt` — new (5.4 validator output for the revised edition)
- `ADVERSARIAL_REVIEW.md`, `adversarial-findings.json` — new
- `sol61-original/` — pre-edit copies of `h-pylori-peptic-ulcer.review.md`,
  `h-pylori-peptic-ulcer.scorecard.json`, `validation-report.json`, `consensus-validation.json`,
  `consensus-addendum.md`, `SHA256.txt`

`h-pylori-peptic-ulcer.scorecard.json` was **not** changed: no level, confidence, count or role moved, and both
corrections are prose-precision fixes. C-1's wording does appear in the scorecard's claim register as well; the lead
may mirror it or leave it, since the figure it carries is unaffected.

## Package-level propagation needed from the lead

No headline string, letter, confidence, count or title changed, so no package-level file carries a now-stale value.
The case folder's `SHA256.txt` does need the new review digest:
`9d3349c208ed2682270bc61b79f93f822b4fbd578f93d57400cb6a0d10539c46` →
`2a75e06a3b6ba9d1c65f2c469ff34b7db7f94c3f13726ad5e67976bd066fa9a6`.

No `propagate_to` or `divergent_copies` entries exist for this slug.

## Lead-flagged Oct-9 revision artifact — checked, not present

The lead reported that another group found a systematic regression in the `consensus_consistency_audit_2026-10-09/revised/`
edition: a spliced Consensus section coexisting with preserved REVIEWER NOTES still asserting that no Consensus work
was performed. **It is not present in this case**, and nothing was edited to "fix" it.

- The spliced block here is headed "Consensus provenance reconciliation — 9 October 2026" and is self-limiting: it
  says the located records are historical checks, "not newly credited complete-paper examinations", that no
  source-comparison count is inferred, and that the supplement "is not a new independent or blinded adjudication".
- The preserved reviewer notes and the scorecard's `literature_tools_note` make **no** negative claim about Consensus.
  They positively enumerate the retrieval tools used by the authoring passes, which predate the 2026-10-05 enrichment
  the Consensus records come from, so the two blocks describe different passes and do not contradict each other.
- Recorded in the review as rejected finding **R-9** rather than silently ignored, per the brief.

Per the lead's instruction, `review_metadata.reviewer` reading "Codex" was treated as house provenance and left
untouched; this reviewer's banner and approval section state the authoring model of record as Sol 6.1 and record the
package's own string verbatim rather than editing it.

## Lead-flagged systematic artifacts — per-case disposition

The lead reported four systematic artifacts found by other groups. All four were checked against this case.

| Artifact | Present here? | Action |
|---|---|---|
| **(1) False approval boilerplate** — the Oct-9 `integrate_consensus.py` appends "The earlier adversarial approval refers to its original edition" unconditionally | **Yes** | Corrected as finding **C-3** in both the review and `consensus-addendum.md`, by rescoping rather than deleting. Verified first: this slug is not one of the 26 genuinely approved on 2026-10-06. |
| **(2) Confidence band justified by reviewer access depth** | No | Recorded as rejected finding **R-11**. This case's High band already rests purely on replication across investigators, countries, ulcer sites, endpoints and designs. |
| **(3) Renderer bug** — "Best-supported rival" printing `models_evaluated[0]`, often the nominal | No (near miss) | Recorded as **R-10**. This is the one case where `models_evaluated[0]` really is the nominal, so the bug could have fired — but the prose correctly names the rank-2 rival M5. |
| **(4) Reviewer-notes contradiction and false provenance clauses** | No | Recorded as rejected finding **R-9**: the spliced block is self-limiting and no preserved sentence contradicts it. The false "existing HIV draft" clause is absent from this batch entirely. |

Per the lead, the scorecard `review_metadata.reviewer` string was treated as house provenance and left untouched.

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

- **Confidence two-step test:** step 1 (strike access depth as a determinant) applied; step 2 (re-derive) -> **High stands**, band unchanged.
- **Momentum stricter rule** ("Decisive work underway" requires work actually being run; a completed publication never satisfies it): **holds**.
- Step 1 changed nothing here - this rationale never leaned on access depth (R-11). Momentum is "Little new work", argued *from* completed work rather than crediting it.

Full derivation is recorded in the review's `## ADVERSARIAL REVIEW AND APPROVAL` section and in `adversarial-findings.json` -> `lead_policy_2026_10_14`.

## Consensus question for the lead

In randomized trials of H. pylori eradication versus control among patients **starting** long-term aspirin or NSAID
therapy, is the reduction in ulcer bleeding sustained beyond about three years in participants with documented durable
clearance? A yes closes the review's only unresolved exception and lifts SC3 toward Θ; a replicated no returns SC1 to Η.