# Adversarial review log — gut-microbiome-systemic-causality (revised edition)

Authoring model: **Sol 6.1** (this edition is a red-team synthesis over two Sol 6.1 / Codex gradings, performed 2026-09-27).
Adversarial reviewer for this round: **Claude Opus 5.5 (Anthropic)**. Date: **2026-10-14**.
Round 2 of the adversarial pass (brief: `deliverables/ADVERSARIAL_REVIEW_BRIEF_2026-10-06.md`, ROUND 2 section).

| Item | Value |
|---|---|
| Level before | Γ — Mechanistically plausible, partial system support |
| Level after | Γ — unchanged |
| Confidence | Moderate — unchanged |
| Role | reinforcing / stabilizing factor — unchanged |
| M / S / D | Mechanistic only / Stable / **Severe overstatement (+3) → Contested** |
| Subclaims | SC1 Β, SC2 Ζ, SC3 Ε (out of scope), SC4 Ε, SC5 Γ — all unchanged |
| PMIDs checked | 38 / 38 resolve; 1 further row is deliberately DOI-only (reference 17, not PubMed-indexed) |
| Claims | 39 registered; **38 verified / 1 unverified** (was 39/0) |
| Provenance gate | **passed → not passed** — the validator requires `gate_passed: false` while any claim is unresolved |
| PMIDs failed | 0 |
| Retractions / EoC | none found; every ledger row carries an explicit `none_found` flag with a check date |
| Load-bearing claims spot-checked | 14 |
| Findings | 0 Critical, 2 Major, 0 Minor, 1 Editorial |
| Validator | PASS — 0 fail, 0 warn, 33 pass under the v5.4 validator this edition declares |

## Findings

| ID | Severity | Location | Change |
|---|---|---|---|
| F1 | Major | Provenance header; crosswalk row C35; AT A GLANCE table and bullet; "Scorecard detail" Divergence row; the two-gradings disagreement table; the Divergence paragraph; score card `provenance_header`, `verification_summary`, `claims[C35]`, `extended_scoring_MSD.divergence` | Claim C35 reclassified **verified → unverified (disputed on the accessible record)**; header counts 39/0 → 38/1; `gate_passed` true → false; Divergence **“Severe overstatement”, implied Ζ, gap +3 → “Contested”** with implied level and gap null. Lead-verified at my request after I flagged the item. C35 paraphrases PMID 40033063 as reporting that the evidence “strongly supported” a causal role in IBD, obesity, T2D and liver disease, and was marked “verified — PubMed abstract or publisher record”; the abstract instead describes a Delphi survey of the **utility of preclinical models**, says disruption “is implicated in” various conditions, and states that establishing cause and effect “is extremely difficult”. The full text is paywalled with no PMC copy, so this pass does **not** assert the source says the opposite — only that the reading is not checkable. C35 was the sole basis for the +3 gap, and the one other field-facing source [34] frames the field at or below the assessed Γ, leaving two sources pointing opposite ways and no single implied level, which is the framework's definition of Contested. Γ and Moderate stand; Divergence modifies no letter. |
| D1 | Major | "Calibration, uncertainty, and evidence quality"; "Reviewer notes" red-team record | "twenty-seven of thirty-nine rows were read at abstract depth" → "thirty-four of thirty-nine", in both places. The sentence names the ledger as its own warrant, and the ledger (and the score card's `citation_ledger`) record 34 `abstract_metadata`, 4 `full_text_partial`, 1 `full_text_complete`. The stated figure understated abstract-only depth by seven rows. Disclosure only — no score input moves. |
| D2 | Editorial | Reference 16 and its ledger row; score card `citation_ledger` and `evidence_inventory` | "…in cirrhosis: THEMATIC trial" → "…in cirrhosis: The THEMATIC trial" (two places in each artifact), matching the PubMed title verbatim. |

## Systematic artifacts reported by the lead — all four checked

1. **False "earlier adversarial approval" boilerplate** — **present but true here, so not corrected.** The clause reads "any earlier approval remain historical records; the earlier approval does not certify this supplement". An earlier independent adversarial pass does exist for this case: the 27 September 2026 red-team synthesis by Claude (Anthropic), recorded in this document's own header and in `reviews/gut-microbiome-systemic-causality/redteam/`. Rejected rather than rescoped.
2. **Confidence set from access depth** — **absent.** The score card's `confidence_rationale` cites only evidence properties (multiple causal animal lines, randomized CDI trials, conflicting human systemic interventions, broad primacy untested). The band was independently re-derived against the High text and Moderate is correct: the review documents demonstrated reproducibility failures rather than well-established reproducibility.
3. **"Best-supported rival" renderer bug** — **absent.** `models_evaluated[0]` is "Host-, disease-, and exposure-first reciprocal ecosystem" with `rank_in_tournament: 1` and `is_nominal: false`; the nominal is the third entry. The prose names the correct rank-1 rival.
4. **Reviewer-notes / Consensus-splice contradiction** — **absent.** This document contains no assertion that no Consensus work was performed, and the spliced block states its own limits ("does not establish historical Consensus use"). No scoping edit was needed.

Score card `review_metadata.reviewer` left as issued; corpus-wide house provenance.

## Revision-regression check

The 9 October revision spliced in the "Consensus consistency validation — 9 October 2026" block. No regression was introduced by that splice: it is self-limiting, it contradicts nothing in the preserved text, and it leaves the grade, scope, ledger and subclaims untouched. The D1 count error predates the splice — it originates in the 27 September red-team record.

## Rejected findings (checked, no change)

1. **Headline Γ** — the strongest-argued letter in group 8 and I could not move it. The Δ steelman is put at full strength and rejected on three named blockers (cap 4 on the human carnitine-withdrawal endpoint, cap 2 on the same report's observational clinical endpoint, and the absence of any pre-onset human perturbation of a prespecified community feature, which cap 5 then blocks forced-transfer evidence from filling by inference).
2. **Scope-matched aggregation** — both Ε subclaims and the Ζ subclaim are explicitly scope-restricted; the declared alternatives (weakest-link Β, primary-subclaim Γ) are correct, and both input gradings pre-registered the same rule.
3. **SC2's Ζ** — four rodent lines named with the distinct systematic error each excludes; Η refused on its own quoted requirement using three named reproducibility failures.
4. **Recurrent *C. difficile* at Ε but excluded from aggregation** — this is the index's own instruction followed, not evaded.
5. **Momentum "Mechanistic only"** — artifact names the work and says why it is not "Decisive work underway".
6. **Divergence +3** — originally accepted because the implied level and arithmetic were both named. **That acceptance is withdrawn:** the lead retrieved the underlying record and the implied level rests on an unverifiable paraphrase. Now finding F1 above.
7. **"Decisive-perturbation studies: 32" vs "28 intervention or perturbation records"** — two different, separately defined quantities (32 = ledger rows flagged `decisive_perturbation`; 28 = a narrower recount whose 18/8/2 components sum correctly). Not a contradiction, but the lead may wish to harmonise one label.
8. **"All 42 ledger rows across both inputs" and "all 36 registered claims"** — input-level totals before the union was reconciled to this edition's 39 rows and 39 claims; the same figures appear in the external process log. Not an error.
9. **Reference 17 DOI-only** — the reference preamble discloses the convention, two targeted PubMed searches returned nothing, and the ledger row records publisher full-text access.
10. **Collective-author reference defect** — checked specifically; reference 35's consortium author renders correctly with no dropped author and no leading comma.
11. **Rendering defects** — none. This edition was hand-assembled rather than produced by the reference builder.

## Residual limitations

Two load-bearing items could not be re-verified here. The Divergence artifact rested on a quotation from the paywalled 2025 Human Microbiome Action Consortium Delphi consensus statement (PMID 40033063) — that the evidence "strongly supported" a causal role in inflammatory bowel disease, obesity, type 2 diabetes and liver diseases. That disease-specific sentence is not in the PubMed abstract, which instead stresses that establishing cause and effect is extremely difficult; the record has no PMC copy and the publisher text was unreachable in this pass. The ledger declares `full_text_partial` access and the review quotes a second, more cautious sentence from the same source, so the attribution was supported by declaration alone. The lead has since confirmed that the accessible record does not support it, which is finding F1: C35 is unverified and Divergence is Contested. A reader with institutional access to the paywalled statement could still settle what it says. Divergence modifies no letter. Second, reference 17, the 2026 active-comparator minimal-hepatic-encephalopathy trial supplying the "68.3% vs 66.1%; p=.80" contrast, is not PubMed-indexed and could not be opened. The external trail this edition cites (`PROCESS.md`, `redteam/`) is not in the deliverables package; it exists at `reviews/gut-microbiome-systemic-causality/`, and the same wrong "27 of 39" figure appears there and should be corrected by the lead.

## Files changed in this case folder

- `gut-microbiome-systemic-causality.review.md` (banner + ADVERSARIAL REVIEW AND APPROVAL section + D1–D2)
- `gut-microbiome-systemic-causality.scorecard.json` (D2, F1)
- `validator-adversarial-2026-10-14.txt` (new)
- `SHA256.txt` (refreshed)
- `sol61-original/` holds the pre-review copies of the review, score card, SHA256.txt and validation-report.json.

Package-level files (`README.md`, `COMPLETION.json`, `FINAL_VERIFICATION.json`) were **not** edited.
Outside this case folder and **not** edited: `reviews/gut-microbiome-systemic-causality/PROCESS.md` (carries the same "27 of 39" error).

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

## Where else the disputed C35 paraphrase appears

Checked at the lead's request. The inventory lists **no divergent copy** of this slug inside `deliverables`, and the phrase does **not** appear in `reviews/gut-microbiome-systemic-causality/PROCESS.md`. It does appear, unchanged, in six places outside this case folder, none of which I edited:

- `the-hypothesis-critic-website/records/gut-microbiome-systemic-causality/` — review **and** score card. **This is the published copy and is the one that matters most.**
- `reviews/gut-microbiome-systemic-causality/final/` — review and score card.
- `reviews/gut-microbiome-systemic-causality/superseded/before-consensus-consistency-2026-10-09/` — review and score card (historical; arguably leave).
- `reviews/gut-microbiome-systemic-causality/redteam/redteam-synthesis-2026-09-27/claim-audit.md` — the row that recorded C35 as verified.
- `reviews/RedTeamRewritten/gut-microbiome-systemic-causality/final/` — review, score card and `claim-audit.md`.
- `sol61-original/` in this folder — the pre-review backup, which must keep the original wording.