# Adversarial review -- gut-brain-psychiatric (ROUND 2)

Authoring model: Sol 6.1. Reviewer: Claude Opus 5.5 (Anthropic). Date: 2026-10-14.
Level before: Γ / Moderate / reinforcing / stabilizing factor. Level after: unchanged.
Findings: 0 Critical, 1 Major, 1 Minor, 1 Editorial. PMIDs verified 23/23. Clinical results spot-checked: 15.
Validator: PASS (34 checks, 0 FAIL, 0 WARN) with prose supplied.

## ADVERSARIAL REVIEW AND APPROVAL


**Authoring model:** Sol 6.1. **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-10-14.

**Scope of checks.** All 23 reference PMIDs verified against PubMed (23/23 resolve to the stated title, first author, journal and year; no retractions or expressions of concern; no empty DOI fields). The reference list was additionally scanned for both forms of the collective-author defect -- a dropped author leaving a stray leading comma, and a consortium name truncated to its last word so that it reads as a surname (Group, Network, Consortium, Investigators, Collaboration, Trial, Study) -- and for dropped nobiliary particles; none is present, and no PubMed record in this ledger carries a collective first author. The eight historical identifiers named in the Consensus provenance block were checked and the three linked artifacts exist at the stated paths. Because this review bears on mental-health treatment decisions, every probiotic and FMT result was checked against its source, including which outcome was prespecified primary and whether the reported analysis was intention-to-treat, modified-ITT or completers-only: PMID 42309058 (Li FMT), 41785480 (Shekarriz allogenic versus autologous), 36637229 (Green pilot), 38732635 (PRO-DEMET, primary F(1,92)=0.58, P=0.45, with the positive minimal-important-difference result secondary), 39504580 (Borkent, 131 randomized, BPRS and BACS primary, null on psychiatric symptoms, zonulin improved, verbal memory borderline at adjusted P=0.059), 28483500 (Pinto-Sanchez, 14/22 versus 7/22, P=0.04), 37314797 (Nikolova, 49 analysed ITT), 35654766 (Schaub, completer sample N=21/26), 40298641 (Moerkl, 43 patients and 43 controls), 28068788 (Romijn, 79 participants analysed as intent-to-treat, null on every psychological outcome), 29731182 (Kazemi, BDI primary, 81 of 110 completers, P=0.042), 40177327 (Elahinejad, 44 of 50 completers, P=0.001, cortisol and ACTH null), 27491067, 37131071 and 27067014 (rodent transfer and vagotomy). Scoring logic re-derived against `hsef-hypothesis-review-v5/SKILL.md` Step 3 and `references/hsef-framework.md`. No level, confidence or role change.

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| G-1 | Major | Consensus provenance reconciliation (9 October 2026) | "The earlier adversarial approval refers to its original edition" becomes "No earlier adversarial approval exists for this case" | No adversarial approval section, banner or approval line exists for this case anywhere in the deliverables tree. The sentence asserted an approval that was never issued, and was introduced by the 9 October revision -- a regression relative to the earlier edition. |
| G-2 | Editorial | Human controlled counterevidence and the pathway gap | "did not outperform autologous FMT on MADRS (16.74 versus 15.4 points, P=.64)" becomes an explicit statement that these are baseline-to-week-24 improvements, with the confidence interval added | PMID 41785480 reports a 16.74-point improvement with allogenic and a 15.4-point improvement with autologous FMT (t = -0.47, P = 0.64, 95% CI -7.3 to 4.6) from baseline to week 24. As printed, the two numbers could be read as endpoint MADRS scores rather than as change scores, which would invert the apparent severity of the sample. The finding, its direction and its role in the argument are unchanged; this is the ledger's single most load-bearing negative result and it should not be misreadable. |
| G-3 | Minor | Human controlled counterevidence and the pathway gap | The Mörkl result is restated as published: the abstract's comparison is probiotic-treated depressed patients **versus healthy controls**, and the randomised within-patient probiotic-versus-placebo contrast is cited from the body (morning RMSSD U = 241.00, Z = 3.03, P = .002; morning lnHF t(33) = −3.48, P = .001, d = −1.19), together with the trial's prespecified primary endpoint | PMID 40298641 abstract: “After 3 months, MD patients receiving probiotics demonstrated significantly improved morning VN function compared to HC.” A treated-group-versus-healthy-control comparison confounds the intervention with the group difference and cannot establish a treatment effect, so describing it as the probiotic arm's result implied a randomised contrast the abstract does not report there. PMC12045568 shows the randomised within-MD contrasts do exist and are significant, so the review's inference survives — but it now rests on the body result that supports it rather than on the abstract comparison that does not. The same source also fixes the endpoint hierarchy (vagal function primary, depressive symptoms secondary), which strengthens the dissociation the review draws. The finding is not load-bearing for a subclaim level: SC4's Γ rests on the absence of any independently perturbed human mediator and on the allogenic-versus-autologous contrast, not on this trial. |

**Rejected findings (checked, no change).**

1. **Probiotic and FMT efficacy statements.** Every intervention result matches its source, including the primary-versus-secondary distinction the brief asks about. PRO-DEMET's null primary comparison and its positive secondary minimal-important-change threshold are both reported, with the review stating explicitly that a secondary responder threshold cannot supersede the prespecified primary. Borkent's null primary psychiatric outcome alongside changed gut-barrier markers is reported exactly. Romijn is correctly described as intent-to-treat and null. Kazemi (81 of 110 completers) and Elahinejad (44 of 50 completers) are both labelled as completer analyses rather than ITT. Schaub's positive time-by-arm interaction is correctly attributed to the compliant subset rather than to the ITT set -- the published abstract reports only the completer sample, and the review's distinction is the stricter reading. Pinto-Sanchez, Nikolova, Green and Moerkl are all reported as their sources state. No change.
2. **The Li FMT trial.** The abstract confirms the two facts the review turns on: remission rates at week 8 did not differ significantly between groups, while HAMD-17 reductions were greater with FMT at weeks 2 and 8. The review's reading -- null prespecified primary, positive secondary -- is therefore correct on the published record. The body-level figures (43.3% versus 26.7%, P=0.1760; -2.9 points, 95% CI -5.5 to -0.3, P=0.0303; roughly 16% of species newly acquired at week 2 falling to about 7% at week 8) could not be independently re-verified here because the article is paywalled and has no PubMed Central deposit; they rest on the review's own recorded full-text read. The review's statement that the abstract's description of engraftment as durable is not what the engraftment analysis shows is a disagreement with the abstract that the review raises openly and attributes, which is the right handling rather than an error. Recorded in residual limitations.
3. **Confidence band.** Checked against the lead policy of 2026-10-14. This review's `confidence_rationale` already rests entirely on evidence density and reproducibility -- small trials, differing products, prespecified primaries disagreeing with the pooled estimates, no human mediator experiment, and key negatives replicated across products, diagnoses and delivery routes -- and makes no appeal to access depth. Moderate stands and no rewrite was needed.
4. **Scope-matched aggregation and the departure from the index rule.** Re-derived. SC4 is the subclaim whose scope matches the graded conjunction and it is Gamma; SC3's higher level sits at a narrower scope that drops the signalling clause and is correctly reported separately rather than setting the headline; SC2's level is restricted to rodents. All three alternative rules are reported, including the index-prescribed weakest-link answer, which is carried in the text as explicit dissent rather than suppressed. The departure is argued on pitfall 21 and on a necessity analysis, not on the letter it produces. Defensible and fully disclosed; no change.
5. **"Best-supported rival" renderer artifact.** Checked the tournament ordering: rank 1 is the context-bounded microbial symptom modifier, which is not this review's nominal model (the nominal M3 ranks 3 of 5). The prose is correct as written.
6. **Momentum "Mechanistic only" and Stability "Stable".** Both carry the required artifact -- named active trial programmes with improving controls, and the explicit absence of an identified near-term trigger. No "Decisive work underway" claim is made, which is correct in any case because the named trials are completed rather than running.
7. **Collective-author defects, both variants.** Scanned and absent; see scope of checks.
8. **Rendering defects.** Scanned for glued numerals, empty `DOI: .` fields and "PMIDnot supplied" strings; none present.
9. **Lossy PubMed abstract fields.** The PubMed abstract of PMID 40298641 is itself garbled mid-sentence — it reads “a significant increase in , particularly along with improved sleep”, with the taxon name dropped (the full text gives *Christensenellales*, particularly *Akkermansia muciniphila*). This is another instance of the lossy-field pattern behind the dropped-collective-author defect, recorded here so the corpus has the example. No text in this review depends on the dropped fragment; the correction it did require is finding G-3.

**Residual limitations.** Seven of the 23 records were read at abstract depth and one (Kazemi) is abstract-only, as the review states; the Mörkl figures added under finding G-3 were taken from the open-access full text (PMC12045568) during this pass and are the only body-level numbers this audit itself introduced. The Li trial's body-level endpoint, engraftment and mediation figures could not be re-verified independently because the article is paywalled; they are the single most load-bearing body-level numbers in this review. This pass is a source, claim and scoring-logic audit, not a blinded recalibration, and nothing in it constitutes clinical advice.

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