# Adversarial review log - akkermansia-metabolic-health (shipped review)

Canonical file: `akkermansia-metabolic-health.review.md` (package `hsef_v53_traffic_ranks_11_15_2026-10-13`).
This log records a second, independent adversarial pass over the shipped review, after the 2026-10-14 pass whose
findings (A1-A4) are printed inside the review itself.

Level before -> after: Γ -> Γ (unchanged). Confidence Moderate -> Moderate (unchanged; no re-derivation changed the band).
Findings: 1 Major, 4 Minor, 2 Editorial. PMIDs verified 59/59 (plus the excluded conceptual record, 60/60 resolved). Claims spot-checked 15. Prior validator: PASS (0 fail, 0 warn, 30 pass); not re-run in this pass.

**Carried-forward item resolved.** The 2026-10-14 pass recorded as a residual limitation that the expression of concern
attributed to PMID23899653 "could not be confirmed through the available PubMed metadata route". It is now confirmed:
PubMed holds an Expression of Concern, PMID 37334997, *Annals of Nutrition & Metabolism* 2023;79(2):107,
DOI 10.1159/000529287 (publication types: Editorial; Expression of Concern), and the score card's
`evidence_inventory` entry 25 carries `ExpressionOfConcernIn: PMID37334997`. The caution printed in REVIEWER NOTES and on
source card [25] is correct as written. See finding A10 for the one place the notice is missing.

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1. **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-11-02.

**Scope of checks.** All 59 REFERENCES PMIDs were re-verified against PubMed `get_article_metadata`: 59/59 resolve to the
stated first author, journal and DOI, 0 failures; the excluded conceptual record PMID41503894 also resolves (Kim HS,
*Gut Microbes* 2026, DOI 10.1080/19490976.2025.2612582, title field empty in PubMed). One editorial notice exists in the
reference list and is the expected one (reference 25, expression of concern PMID 37334997); no retraction and no other
notice was found. Fifteen load-bearing claims were spot-checked against the cited abstracts and, for the pilot, against the
PMC full text (PMC6699990), concentrating on the records that hold or raise a letter: the 2019 MucT pilot (40 enrolled /
32 completed; 11 placebo, 12 pasteurized, 9 live; significantly higher baseline insulinemia and lower insulin sensitivity in
the pasteurized arm, P = 0.014 and P = 0.007 versus placebo; insulin sensitivity +28.62 +/- 7.02%, P = 0.002; insulinemia
-34.08%, P = 0.006; total cholesterol -8.68%, P = 0.02; body weight -2.27 kg, P = 0.091; "Fasting glycemia was not affected"
and "HbA1c was not modified by the supplementation", with the stated non-hyperglycemic, non-diabetic baseline); the 2026
maintenance trial (n = 90, 8-week low-energy diet for >= 8% loss, 24-week maintenance, regain 1.2 +/- 0.7 kg versus
3.2 +/- 0.4 kg, P = 0.012); the 142-person multicenter trial (4 months, Matsuda primary not different in the
intention-to-treat set, exploratory low-baseline benefits, weight P = 0.06, trunk fat P < 0.05); the HB05 trial (100 adults
aged 60+, 12 weeks, left-leg extensor peak torque P = 0.0103 and peak torque/body weight P = 0.0052, follistatin P = 0.0063);
the AKK-WST01 trial (58 adults, 12 weeks, HbA1c and weight fell in both arms with no significant between-group difference,
benefit and colonization confined to low baseline); the WBF-011 trial (n = 76, glucose total AUC -36.1 mg/dL/180 min,
P = 0.0500, A1c -0.6); Plovier (pasteurization enhanced the effect; Amuc_1100 interacts with TLR2 and partly recapitulates);
Everard 2013 (live reversed the high-fat-diet phenotype, heat-killed cells did not); the inulin pilot (body weight -2 kg and
lower insulin requirement, HbA1c and CGM unchanged, no rise in *Akkermansia*); Allin (faecal transfer from prediabetic donors
did not impair glucose regulation in recipient mice); Van Hul (*Subdoligranulum variabile* supplementation produced no
significant difference in any obesity or diabetes hallmark); the postbiotic-yogurt trial (waist, waist-to-height, body fat
per cent, AST and appetite versus control; no physical-fitness contrast reported); Hong (heat-killed EB-AMDK19 suppressed
weight and fat gain in beagles); Kumar (novel *Akkermansia* species, live form more pronounced); and Choi (pasteurized more
potent than live for glucose tolerance). Every checked number, species, system and direction matched its source. A targeted
search for contradictory or omitted literature on the registered claim returned no further human pasteurized-MucT trial and
no MucT meta-analysis, so the clinical evidence base for SC2 and SC5 is complete as cited. Scoring logic was re-derived from
the inventory: caps precede aggregation; `boundary_check_BGD` gives ceiling Γ for SC1-SC4 and Δ for SC5; the declared
`primary-subclaim` rule returns SC2 = Γ, which matches the printed headline, and the printed sensitivities
(weakest-link Α via SC4, primary-subclaim Γ, scope-matched Γ) are consistent with the matrix; the tally 1 positive /
42 screened discriminating contrasts matches `causal_tally_audit` and the 42 inventory records flagged
`discriminating_contrast`, and the eight printed inventory entries ([1], [2], [3], [4], [6], [39], [40], [43]) are exactly the
eight records flagged `decisive_perturbation`; the two falsification patterns are met with no downgrade applied, as printed;
and M/S/D does not alter the level (`role_of_mechanism.tournament_used_as_tiebreaker: false`, step 5 label derived from causal
direction). The competing-model tournament was checked for fairness: all four models carry named supporting and contradictory
evidence and the dietary and strain/processing alternatives are real rivals rather than strawmen, but see A7 on how their
ranks are presented.

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| A5 | Major | `reader-summary.json`, `limitations` (reader-facing) | "59 primary reports with 2 selected partial-body examinations; density, heterogeneous controls and incomplete access set the confidence, not correctness." to "59 primary reports with 2 selected partial-body examinations; evidence density, heterogeneous controls and reproducibility set the confidence, not correctness or access depth." | This is the exact boilerplate the 2026-10-14 pass struck from the score card under lead policy (finding A4), on the ground that access depth is not evidence density and never justifies a band. The score card now reads "Moderate reflects evidence density and reproducibility, not access depth", so the reader-facing summary states the opposite basis for the same band. Of the two, the reader summary is the lay-facing text. |
| A6 | Minor | `reader-summary.json`, `redteamReviewed` | `false` to `true` | The review banner states "adversarially reviewed by Claude Opus 5.5 (Anthropic) on 2026-10-14. 4 corrections applied", the review carries an ADVERSARIAL REVIEW AND APPROVAL section with findings A1-A4, and this second pass exists. The flag contradicts the shipped document. |
| A7 | Minor | COMPETING MECHANISTIC MODELS, closing sentence; AT A GLANCE, "Best-supported rival" | Add, after "The tournament does not alter the nominal level": "Ranks measure how well each model accounts for the evidence bearing on the registered claim, not each model's own HSEF level; rank 3 carries Δ for its own dietary scope while ranking below two Γ models on the registered claim." | Rank 3 ("Diet and baseline host/community state drive response") is printed with level Δ, a higher letter than the Γ carried by the nominal model and by the rank-1 "best-supported rival", yet it is ranked third and is not the named rival. Score card `competing_models.models_evaluated` confirms M3 `rank_in_tournament: 3`, `hsef_level_for_this_model: "Δ"`. Nothing printed tells a reader that rank and level are different axes, so the tournament reads as internally inconsistent. |
| A8 | Minor | Upward threshold ladder, SC5 Δ rung | "SC5 Δ: “Requires at least one intervention changing lifespan or healthspan in a way consistent with the hypothesis.” Satisfied: yes. " (line ends; no text follows) to the same line followed by the crossing evidence, e.g. "Satisfied: yes. The 90-person randomized maintenance trial (PMID42120725) reports 1.2 versus 3.2 kg regain over 24 weeks after an eight-week diet lead-in; the crossing is scoped to that maintenance setting and is unreplicated." | Every other rung in the ladder prints a next-step or evidence sentence after the satisfied verdict; this one is empty. The single rung in the whole review where a Δ is actually crossed therefore names no study, no effect size and no scope. Verified in the file: the line terminates after "Satisfied: yes." with a trailing space. |
| A9 | Minor | COMPETING MECHANISTIC MODELS rank 4; SC3 cap audit | "Early heat-killed failure, later pasteurized improvements" and "early heat-killed cells fail" to wording that scopes the failure to the single 2013 preparation, e.g. "the one heat-killed preparation tested in 2013 failed, while later non-viable preparations succeeded in other species" | Two of the review's own cited records report positive non-viable preparations: [7] heat-killed EB-AMDK19 suppressed weight gain, fat deposition and triglyceride rise in beagles over 12 weeks and its abstract states the heat-killed strain "exerts similar protective effects as the live bacterium in mice" (PMID 39264284, DOI 10.1080/1745039X.2024.2397221); [44] reports heat-killed *Akkermansia* improving metabolic and muscle outcomes in STZ mice (PMID 39429872). Both source cards state this correctly, so the defect is that the model summary and the cap rationale generalise from the 2013 result and are contradicted by the inventory beneath them. The letter does not change: SC3 stays Γ either way. |
| A10 | Editorial | REFERENCES, entry 25 | Add the notice to the printed reference, e.g. "... Access: abstract_metadata. Study: PMID23899653. Editorial notice: expression of concern, PMID 37334997." | The expression of concern appears in REVIEWER NOTES, on source card [25] and in the score card's `status_flags`, but not on the reference-list line itself, which is where a reader scanning citations looks. Confirmed notice: PMID 37334997, *Ann Nutr Metab* 2023;79(2):107, DOI 10.1159/000529287. |
| A11 | Editorial | PROPOSED DECISIVE EXPERIMENTS; decisive-perturbation tally sentence | Four doubled full stops ("... mechanistic leads.. Controls: ... analyses, not selection of outcome-defined responders.. Supporting: ... evidence.. Falsifying: ... results..") to single stops; "Tally: 1 positive qualifying study units" to "Tally: 1 positive qualifying study unit" | Renderer artefact and number/noun disagreement. No content added or removed. |

**Rejected findings (checked, no change).**
(i) Reference years printed as print or issue years that differ from the PubMed date - references 6, 12, 28, 33, 35, 37, 38,
42 and 55. Each was checked against the journal citation and is correct as printed: reference 28 is *Gut Microbes* 17(1):2438823
(e-pub 2024-12-22, 2025 issue), reference 35 is *J Acad Nutr Diet* 120(4):650-659 (e-pub 2018, 2020 issue), reference 6 is
*Nat Med* 23(1):107-113 (e-pub 2016-11-28, 2017 issue). Declared non-error.
(ii) PubMed title fields that drop italicised species or strain names - "PasteurizedHB05" at reference 3, "Muc" standing in
for the strain name throughout the Mount abstract, and reference 48's title field reduced to a single full stop. The review is
more accurate than the metadata in each case; declared non-error.
(iii) Template bleed: searched for text from unrelated hypotheses. The strings "NAD", "ROS", "senescence", "mitochondrial"
and "lifespan/healthspan" appear only inside quoted rubric 06 criteria and their explicit "Not triggered literally"
dispositions, which the review labels as quotations and pairs with a stated adaptation. No sentence from an unrelated case
was found anywhere in the file, and the body records "Lifespan is not imposed on the non-lifespan treatment claim."
(iv) Rubric 06's literal lifespan criterion quoted at the SC2 and SC5 Δ rungs: same disposition as the 2026-10-14 pass -
rewriting a quoted framework criterion would be the larger error. Recorded again as a residual limitation.
(v) Tally arithmetic: "1 positive qualifying study units ... 42 screened discriminating contrast units" matches
`causal_tally_audit.tally_after_audit` = 1 and `screened_discriminating_contrasts` = 42, and 42 inventory records carry
`discriminating_contrast: true`. The eight printed inventory entries are the decisive perturbations (8 records flagged), a
different set; the review does not conflate them.
(vi) SC4 at Α and the weakest-link sensitivity of Α: SC4 is a deliberately stated unsupported inference
(generic abundance or class effect to named product), which makes weakest-link Α arithmetically correct. The headline is set by
the preregistered `primary-subclaim` rule, which is declared before the sensitivities are printed, so the Α sensitivity is
not load-bearing and is not presented as the grade.
(vii) Momentum "Mechanistic only": correct. No named pending trial is asserted, and the completed 2026 trials are described as
completed.
(viii) No collective-author defect, no empty DOI field, no glued-numeral or fused-word defect was found in the prose of this
review; the only glued strings are inside PubMed-sourced titles (see (ii)).
(ix) Missing contradictory literature: a targeted PubMed search for pasteurized MucT clinical trials returns only PMID42343233
and PMID42120725, both cited, and no systematic review or meta-analysis of *Akkermansia* supplementation in humans was found
that bears on the registered endpoint. Nothing contradictory is omitted.
(x) Stability "Stable": argued both ways and left as published. Two pivotal 2026 trials entered the evidence base inside the
review's own window, one of them discovered mid-session and forcing the post-discovery SC5 at 2026-10-13T16:18 against a
16:06 registration, which reads against "Stable". But the dimension keys on a named pending result, the review states that
none was identified, and the score card records the same justification, so the value is correct as defined.
(xi) Levels: SC1-SC5 and the headline Γ follow from the quoted caps, `boundary_check_BGD` and the declared aggregation rule;
no letter required a change, and M/S/D did not move one.

**Scorecard change.** None proposed. `hsef_level`, `confidence`, subclaim levels, the tally and every other score card field
are unchanged by this pass. The corrections proposed above fall in `reader-summary.json` (A5, A6) and in the review prose
(A7-A9, A10, A11); none was applied by this reviewer.

**Residual limitations.** Two of 59 records were examined at partial-body depth, as the banner states; this pass added the PMC
full text of the 2019 pilot for verification only and claims no complete-paper audit of the remainder. Rubric 06's literal
aging wording remains in the threshold-ladder quotations. Scite verification was unavailable in this pass (provider monthly
limit reached), so retraction and notice screening rests on PubMed publication types and comments-corrections records only.
The framework validator was not re-run, because no file was changed.

**Disposition:** The review's letter, confidence, caps, aggregation, tally and competing-model content survive this pass. One
Major defect is reader-facing and sits outside the review file, in `reader-summary.json`. Correction of A5 through A11 is
recommended before any further publication step; no finding requires a change of level or confidence.
- Claude Opus 5.5 (Anthropic), 2026-11-02.