# Adversarial review log - weight-loss-mash-reversal (shipped review)

Canonical file: `weight-loss-mash-reversal.review.md` (package `hsef_v53_traffic_ranks_11_15_2026-10-13`).
This is a second, independent adversarial pass over the shipped review. The review already carries an in-document
`ADVERSARIAL REVIEW AND APPROVAL` section dated 2026-10-14 (findings WM-1 to WM-5). Findings here are numbered WL-1 onward so
they cannot be confused with that pass. One finding (WL-1) overturns a judgement that pass recorded as a rejected finding.

Level before -> after: Δ -> Δ (unchanged). Confidence Moderate -> Moderate (unchanged; rationale should be rewritten, see WL-1 and WL-2).
Findings: 2 Major, 5 Minor, 1 Editorial. PMIDs verified 60/60. Claims spot-checked 23. Validator: not re-run (no file edited in this pass); last recorded PASS (0 fail, 0 warn, 31 pass) per `validation.txt`.

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1 (per the review's own banner; the score card's `provenance_header.generated_by` reads "Codex; HSEF v5.3"). **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-11-02.

**Edition note.** No correction has been applied by this pass. Every finding below gives the exact original text and a
proposed replacement for the lead to adjudicate. The review, score card, reader summary, site pages and flags were not edited.

**Scope of checks.** All 60 REFERENCES PMIDs re-fetched with PubMed `get_article_metadata`: 60/60 resolve to the stated first
author, journal, title and DOI (PMID 26367736 has no DOI in PubMed, correctly printed as "not indexed in PubMed"); 0 failures;
no retraction or expression-of-concern publication type on any record. Year differences of one at about a dozen records
(for example 27521509, 36549296, 33185364, 26608256, 36508737, 39694202) are e-publication versus print year and acceptable
under the house convention. 23 load-bearing claims were spot-checked: against PMC full text for the two lifestyle trials that
hold the level (Promrat, PMC2799538; Mucinski, PMC12007730) and against PubMed abstracts for Vilar-Gomez, Houghton, Abad (ESG),
Harrison 2004 (orlistat pilot), Lee (balloon), Hansen (LCHF), Cusi (pioglitazone), ESSENCE, the F4 semaglutide trial, BRAVES,
NATIVE, SYNERGY-NASH, Lassailly, LEAN, Cuthbertson, Dallio, Koutouratsas and Takahashi. Every drug, surgery and cohort claim checked
matched its source in direction, system and size. Scoring was re-derived from the review's own inventory: the declared
`primary-subclaim` rule takes SC2 = Δ as the headline; the weakest-link sensitivity min(Δ, Δ, Γ, Γ) = Γ is correctly printed as a
sensitivity only; no universal cap or literal Γ hard rule binds a randomized change in measured biopsy pathology; M/S/D do not
touch the level. The score card holds `decisive_perturbation: true` for exactly the 12 printed inventory records and
`discriminating_contrast: true` for exactly 33, matching the prose. A targeted PubMed search for randomized paired-biopsy
lifestyle trials was run to test for omitted contradictory evidence (WL-2).

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| WL-1 | Major | Evidence and causal interpretation (para. 3); SC2 rationale in the subclaim matrix and SC2 field-valid line; inventory [46] and evidence card [46] "Does not show"; falsification "Unchecked numerical reconstruction"; steelman; Stability; KNOWLEDGE GAPS and Research gaps; REVIEWER NOTES; Consensus section; 2024 timeline entry; score card `final_score.confidence_rationale` | "Results state nine treatment resolutions but only eight ballooning-zero participants, an internally inconsistent exact count that is left unresolved. We retain the reported qualitative clinical positive with Moderate confidence and seek clarification before using a precise resolution fraction." to "Results report resolution of steatohepatitis in 9 treatment and 0 control participants (final sample 16 and 8; the Discussion gives 56%). Eight treated participants *reduced* ballooning to 0; this is compatible with nine resolutions, because inclusion required NAS of at least 4 rather than ballooning of at least 1, so a participant could enter with ballooning 0. The counts are reconcilable, not inconsistent. The real caveats are that component reductions are reported within the treatment arm, the text prints no between-group test for resolution or components, and control participants entered with higher MRS liver fat (24% versus 15%), which the authors say may have contributed to the difference." Make the same change at every other listed location; in the knowledge gap, replace "Source clarification of the nine-resolution/eight-ballooning-zero inconsistency in PMID38914313." with "Baseline ballooning distribution and a printed between-group resolution test for PMID38914313."; strike the discrepancy from the confidence rationale | PMC12007730 Methods: "Sedentary men and women with a NAFLD activity score (NAS) ≥4/8 ... were recruited"; "resolution of MASH corresponds to ... 0–1 for inflammation and 0 for ballooning". Results: "Resolution of steatohepatitis was achieved in 9 treatment subjects and 0 control subjects ... 8 reduced hepatocellular ballooning to a 0". Discussion: "the 56% of participants in the current study who achieved MASH improvement with treatment (0% of controls)". Limitations: control "IHTG ... differed between the groups and may have contributed to greater histologic improvements observed with treatment". The review says no fraction can be given, yet the paper gives one. The supposed inconsistency appears about 12 times and is one of four named determinants of the confidence band, so this misreading of the study that sets the level is Major. The 2026-10-14 pass recorded the inconsistency as "real"; that judgement is overturned. **Level unchanged**: 9/16 against 0/8 still clears Δ (a two-sided Fisher exact test on the printed counts gives p of about 0.01, the reviewer's own arithmetic, not a figure from the paper). **Confidence unchanged**: Moderate still rests on a single small, unreplicated, completer-sample trial with baseline imbalance, plus WL-2. |
| WL-2 | Major | SC2 rationale; decisive-perturbation inventory; tally; M1 tournament entry | Absent to an added inventory item: "[new] Eckard et al. 2013 randomized 56 adults with biopsy-proven NAFLD (88% of completers with NASH) to standard care or one of three diet/exercise arms for six months with repeat biopsy. NAS improved in the 41 completers as a whole (p < 0.001) with no difference between arms, standard care included (p = 0.31). A randomized lifestyle-versus-standard-care paired-biopsy contrast was therefore null; mixed NAFLD/NASH entry, completer analysis and small arms limit it, and it does not refute the positive SC2 trial." Add to the SC2 rationale: "The other randomized standard-care-controlled paired-biopsy lifestyle trial found on targeted search (Eckard 2013) showed improvement in every arm without a between-arm difference, so the assigned-package effect is not yet consistently reproduced." Screened discriminating contrasts 33 to 34 | PMID 23814606 (doi:10.1177/1756283X13484078), verified via PubMed: "56 participants were randomly assigned to 1 of 4 lifestyle modification subgroups for 6 months: standard care, low-fat diet and moderate exercise, moderate-fat/low-processed-carbohydrate diet and moderate exercise, or moderate exercise only ... Among the 41 participants who completed the study (88% with NASH), a significant change was found in pre- to post-NAFLD activity score in the group as a whole (p < 0.001) with no difference detected between subgroups (p = 0.31)." This is the same kind of randomized paired-biopsy assigned-lifestyle contrast as the two trials the review relies on, inside the 24–104-week window, and its result points the other way. The review's own inventory lists only the two positive lifestyle trials. **Level unchanged** (Δ needs one valid positive). It strengthens, not weakens, the Moderate band. |
| WL-3 | Minor | Decisive-perturbation inventory, tally sentence; score card `causal_tally_audit` | "Tally: 1 positive qualifying study unit across scoped subclaims; 33 screened discriminating contrast units." to "Tally: 2 positive qualifying study units across scoped subclaims ([1] for SC1 biopsy steatosis; [46] for SC2 inflammation/ballooning resolution); 33 screened discriminating contrast units (34 with WL-2)." Score card `tally_claimed_by_document`/`tally_after_audit` 1 to 2 | The review crosses SC1 at Δ on [1] ("Positive full-scope result: yes"; `boundary_check_BGD` SC1 `qualifying_evidence_citation: PMID19827166`). Promrat full text confirms the steatosis contrast: "Steatosis score also improved to a significantly greater degree in the lifestyle group as compared to the control group (−1.1 [0.8] vs −0.3 [0.8], p=0.02)". A tally defined "across scoped subclaims" that counts only the SC2 unit leaves the SC1 crossing without a counted unit. Alternatively, keep 1 and re-label the unit as "for the primary SC2". |
| WL-4 | Minor | COMPETING MECHANISTIC MODELS; score card `competing_models` | The four models are all broadly compatible with the claim (component dissociation, mediator confounding, stage transfer). Add a null rival: "Rank 4/5: Non-specific trial effects, regression to the mean and biopsy sampling variability explain small-trial histologic change (Γ). Control and standard-care arms also improve; single-pathologist small samples and completer analyses magnify noise. Randomized between-arm differences in [1] and [46] are not explained by this, but the null between-arm contrast in Eckard 2013 is." | Promrat full text: "histologic parameters of NASH activity ... improved although to a lesser extent even in those who had minimal weight loss or those assigned to the control group ... the improvement in serum ALT could be partly due to the phenomenon of regression to the mean." Mucinski: control participants also lost weight and liver fat ("Both groups exhibited reductions in ... liver fat"). Eckard 2013 (WL-2): improvement in all arms including standard care. Without a model built on the control arms, the field looks fair but omits the most direct alternative to a package-specific effect. |
| WL-5 | Minor | Cap audit: SC1–SC4 "field-valid definition" lines; Upward threshold ladder SC1/SC2 "Δ: ... Satisfied: yes" | "SC2 Δ: “Requires at least one intervention changing lifespan or healthspan in a way consistent with the hypothesis.” Satisfied: yes." to "SC2 Δ: “Requires at least one intervention changing lifespan or healthspan in a way consistent with the hypothesis.” Literal lifespan/healthspan wording does not apply to this clinical claim; satisfied under the declared clinical adaptation (assigned treatment changing paired biopsy pathology): yes." Same change for SC1, and a one-line preface before the four field-valid quotes stating that they are the literal rubric 06 aging text | The ladder prints "Satisfied: yes" directly under a quoted lifespan/healthspan criterion and "Must separate developmental effects from true aging effects", which no liver-histology trial can satisfy literally. The adaptation is declared in "Rubric fit" and is policy-compliant, so this is not template bleed in substance. A lay reader would still see a liver trial certified as extending lifespan. |
| WL-6 | Minor | `reader-summary.json` `limitations` (lay-facing) | "60 primary reports with 3 selected partial-body examinations; density, heterogeneous controls and incomplete access set the confidence, not correctness." to "Moderate confidence rests on a single small randomized trial of current MASH resolution that has not been replicated, a second randomized trial with mixed component results, and heterogeneous packages and comparators; it bounds the evidence base, not correctness." | WM-2 struck access depth from the score card's confidence rationale under lead policy, but the reader summary still carries the struck sentence. The same stale string appears in the reader summaries of at least 30 other slugs (batch-level propagation item; not edited). |
| WL-7 | Minor | Evidence cards [33], [38], [50], Access/status lines | "ErratumIn: PMID41880634" to "Erratum In: PMID41880634"; "ErratumIn: PMID32738952; ErratumIn: PMID34147150" to "Erratum In: PMID32738952; Erratum In: PMID34147150"; "ErratumIn: PMID31386114" to "Erratum In: PMID31386114" | Known renderer defect that drops spaces (same class as the super-enhancer S3 finding). No content change. |
| WL-8 | Editorial | PROPOSED DECISIVE EXPERIMENTS, Readout/Controls/Supporting/Falsifying line | "separate endpoints.. Controls:" / "multiplicity addressed.. Supporting:" / "their separate claims.. Falsifying:" / "achieved-loss responders.." to single full stops | Doubled punctuation from field concatenation. |

**Rejected findings (checked, no change).**
(i) SC2 at Δ rather than Γ: a randomized assigned package changed paired biopsy inflammation/ballooning resolution (9/16 against 0/8),
which is measured disease pathology, not a biomarker. No universal cap or literal Γ hard rule forces Γ. Retained.
(ii) SC1 at Δ: Promrat's between-group steatosis contrast (p = 0.02) and the NAS responder contrast (13/18 against 3/10, p = 0.03) were
confirmed in full text; 28 of 31 had paired biopsies (18/21 lifestyle, 10/10 control), as the review states.
(iii) Ε not reached: there is one current-resolution positive, and WL-2 adds a randomized null. The steelman's rejection of Ε is correct,
and is if anything understated.
(iv) Aggregation: the declared rule is primary-subclaim, applied as written; the weakest-link Γ and scope-matched Δ sensitivities are
correctly labelled as sensitivities.
(v) Best-supported rival label: `models_evaluated[1]` (M2) carries rank 1 and `is_nominal: false`; the AT A GLANCE text, as corrected
by WM-1, names it correctly.
(vi) Momentum "Mechanistic only": ESSENCE's 240-week phase is a drug trial and is correctly not counted as pending lifestyle work.
(vii) Drug, surgery and cohort cards: ESSENCE (1197 randomized, interim 800, 62.9% against 34.3% resolution), F4 semaglutide (71; fibrosis
11% against 29%, p = 0.087), BRAVES (288; 56%/57% against 16%), SYNERGY-NASH (190 randomized, 157 evaluable biopsies), Lassailly
(64 of 94 at five years), LEAN (45 of 52 biopsied), NATIVE (1200 mg significant, 800 mg p = 0.07), pioglitazone (+2.5 kg weight gain),
Koutouratsas (decompensation HR 0.72, MACE HR 0.98) and Takahashi all match their abstracts.
(viii) Orlistat pilot [15] is called "uncontrolled" although PubMed's publication-type field lists "Randomized Controlled Trial". The
abstract describes a ten-patient single-arm study, so the review is correct and the index tag is wrong.
(ix) Harrison et al. 2009 orlistat RCT (PMID 19053049) is not cited. It randomizes orlistat on a shared diet, so it adds only an
achieved-loss association (at least 9% loss with histologic improvement) that is already well represented. The omission is not a defect.
(x) Role of mechanism "reinforcing / stabilizing factor" for a therapeutic package: the derivation explicitly declines endogenous
primacy and is consistent with step 5. Left as written.

**Scorecard change (proposed, not applied).** `final_score.confidence_rationale` (strike the nine/eight discrepancy as a
determinant and add the randomized null of WL-2); `causal_tally_audit` tally 1 to 2 and screened contrasts 33 to 34;
`competing_models` add the null-effects model of WL-4; `evidence_inventory` add PMID 23814606. Level, confidence and subclaim
levels unchanged.

**Residual limitations.** Full text was read for two of 60 records (PMIDs 19827166 and 38914313); the rest were checked at abstract
depth. Mucinski figures and supplementary tables were not available in the retrieved text, so between-group component statistics
could not be confirmed or excluded. The PubMed search for omitted randomized lifestyle paired-biopsy trials was targeted and not
systematic. The provenance mismatch (banner "Sol 6.1" against score card "Codex") is carried forward from the earlier pass and was not
resolved here.

**Approval:** Not approved as is. Approval is conditional on the lead adjudicating WL-1 and WL-2 and on the proposed corrections being applied; no level or confidence change is required. - Claude Opus 5.5 (Anthropic), 2026-11-02.