# Adversarial review log - glp1-microdosing (shipped review)

Canonical file: `glp1-microdosing.review.md` (HSEF v5.3). This is an independent second adversarial pass over the shipped review,
which already carries a first adversarial approval dated 2026-10-14 (findings GM-1 to GM-4). Findings below are new unless stated;
none of them repeats GM-1 to GM-4, and those four corrections were confirmed as applied.

Level before -> after: Γ -> Γ (unchanged). Confidence Moderate -> Moderate (unchanged).
Findings: 1 Major, 8 Minor, 2 Editorial. PMIDs verified 59/59 (plus linked correction PMID 41932351). Claims spot-checked 17. No file other than this log and its JSON was written.

## 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 re-fetched with PubMed `get_article_metadata`: 59/59 resolve to the stated first
author, journal and DOI; 0 failures; no retraction or expression-of-concern flag. Year differences of one at references 7, 22, 27, 29, 51
and 58 are e-pub versus print and are accepted. The linked notice PMID 41932351 resolves to "Correction to Lancet Diabetes
Endocrinol 2025; 13: 935-48" (Published Erratum, STEP UP T2D parent). Seventeen load-bearing claims were spot-checked against PubMed
abstracts: the daily dose-ranging trial [1] (957 randomised, -6.0% to -13.8% at 0.05-0.4 mg daily vs -2.3% placebo, every dose
once daily); weekly dose-finding [22] (415 patients, 12 weeks, 0.1-0.8 mg without escalation, largest weight effect -4.8 kg at 1.6 mg
escalated); the 60-week intake trial [8]; STEP 1 [9]; STEP 2 [34]; STEP 3 [35]; STEP 4 [32]; FLEX-SEMA [2]; INFORM [19]; SEMALEAN [51];
the older-diabetes cohort [52]; the Cureus chart review [16]; the reduced-frequency series [50]; the bimagrumab mouse study [53]; the
PCOS pilot [58]; STEP UP imaging [11]; and the pooled GI analysis [55]. Numbers, species/system and direction matched in every case;
the defects found are of selection, framing and omission, not of misquoted figures. Scoring was re-derived from the score card:
the declared `primary-subclaim` rule returns SC2 = Γ, matching the headline; the weakest-link sensitivity Α follows from SC3 = Α; caps
are quoted before aggregation; `causal_tally_audit` and `evidence_inventory` were cross-checked against the printed inventory; M/S/D
are stated not to alter the level. Targeted searches (PubMed; ClinicalTrials.gov) looked for contrary or on-scope literature published
before the 2026-10-13 cutoff and for named pending trials.

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| A1 | Major | Evidence and causal interpretation; decisive-perturbation inventory; SC2 decision; Momentum | No mention of the largest on-exposure dataset -> add an inventory item: "A federated EHR analysis identified 814 adults with at least three 0.25 mg-only semaglutide prescriptions spanning at least six months; mean 12-month weight change was -2.2% (versus -5.5% on sustained 2.5 mg tirzepatide). Observational, uncontrolled for placebo and selected by prescription pattern, it is the closest real-world record of sustained sub-0.5 mg weekly exposure and does not show clinically meaningful (at least 5%) loss." Add it to the SC2 rationale and the Momentum artifact. | PMID 42857675 (Murugadoss et al., Biol Methods Protoc 2026; DOI 10.1093/biomethods/bpag048; e-pub 2026-08-31, before the cutoff). Verified abstract: "Among 490 072 semaglutide-treated patients, 814 met the stringent sustained low-dose criterion ... mean 12-month percent weight change -5.5% vs. -2.2%". Not in the score card or citation ledger. It is the only sizeable dataset at the registered exposure and it points against the claim; omitting it leaves readers with only favourable small low-dose cohorts ([16], [50], [58]). It does not move SC2 below Γ because it has no placebo arm and is prescription-selected. |
| A2 | Minor | Evidence and causal interpretation, para. 2; Adversarial steelman (above) | "Every semaglutide injection in that trial was daily. Multiplying or dividing its milligrams cannot manufacture a weekly exposure trial." -> "Every semaglutide injection in that trial was daily, and the preregistered scope excludes daily-to-weekly conversion. Pharmacologically, however, a sponsor exposure-response model found that systemic exposure, not schedule, explained weight trajectories across the daily 0.05-0.4 mg and weekly 2.4 mg trials; the 0.05 mg daily arm (0.35 mg per week cumulatively) is therefore the closest exposure-matched randomised evidence, and it is the strongest argument a Δ steelman could make." | PMID 37424165 (Strathe et al., Diabetes Obes Metab 2023; DOI 10.1111/dom.15211): "exposure levels over time consistently explained the weight-loss trajectories across trials and dosing regimens". PMID 30122305 abstract: 0.05 mg daily gave -6.0% vs -2.3% placebo at 52 weeks. The scope exclusion stands, so no level change; but the prose states a pharmacological impossibility that the cited literature contradicts, and the steelman omits its strongest upward argument. |
| A3 | Minor | SC4 decision; Falsification ("Necessary subjective food-noise explanation"); inventory item [8]; Best-supported rival | "In PMID 42323166, intake and weight differences persist at 40/60 weeks while appetite and food-preoccupation contrasts lose significance." -> "In PMID 42323166, intake differences persist at 20/40/60 weeks while hunger and food-preoccupation contrasts lose significance after week 20; food-reward responsiveness (Power of Food Scale), the measure closest to the food-noise construct, still differed at week 40 and only at week 60 did no subjective contrast remain." | PMID 42323166 abstract: "greater reductions in responsiveness to the rewarding value of food at weeks 20 and 40 (MD = -0.4 and -0.6)". The review presents the dissociation as cleaner than reported; the hedonic contrast that persisted is the one nearest to "food noise". SC4 stays Γ. |
| A4 | Minor | Evidence and causal interpretation, para. 4 (last sentence); inventory [32] and [50]; evidence card [50] | "None satisfies the registered initiation claim." -> "None satisfies the registered maintenance claim." (and "do not test below-0.5 mg once-weekly initiation" -> "do not test below-0.5 mg once-weekly maintenance against placebo") | Registered claim and SC2 are about "maintenance doses below 0.5 mg ... sustained weight loss", not initiation; the review's own decisive experiment adds a de-escalation arm "without redefining the initial-treatment claim". The prose alternates between an initiation reading and a maintenance reading, which changes how STEP 4 and the reduced-frequency series should be weighed. Internal contradiction; no level change. |
| A5 | Minor | `reader-summary.json` `limitations` (lay page) | "59 primary reports with 3 selected partial-body examinations; density, heterogeneous controls and incomplete access set the confidence, not correctness." -> "59 primary reports examined. Moderate confidence rests on evidence density and reproducibility, not reading access: no study tests the registered sub-0.5 mg weekly maintenance exposure, the dose-adjacent reports are uncontrolled and response-selected, and muscle and food-noise outcomes conflict or rest on recall." | GM-3 corrected this rationale in the score card under lead policy, but the lay summary still prints the struck wording; prose, score card and reader summary now disagree. |
| A6 | Minor | Decisive-perturbation inventory, tally sentence | "Tally: 0 positive qualifying study units across scoped subclaims; 11 screened discriminating contrast units." -> same counts, plus: "The eleven discriminating contrasts ([22], [30], [32], [34], [50], [51], [52], [53], [54], [55], [58]) are a different set from the twelve decisive perturbations printed above." | Score card `evidence_inventory`: `decisive_perturbation: true` for exactly the 12 printed items ([1],[2],[8],[19],[22],[32],[34],[50],[51],[52],[53],[58]); `discriminating_contrast: true` for 11, including [30], [54], [55] that are not printed and excluding [1], [2], [8], [19]. Both numbers are correct but readers cannot reconcile 11 with 12. Note also that [8], the review's principal falsification source, is not flagged discriminating; the lead may wish to check that flag. |
| A7 | Minor | AT A GLANCE table, "Causal role" cell | "epiphenomenon" -> "epiphenomenon (framework label for no demonstrated dependency at the exact registered dose; standard-dose semaglutide is causally effective)" | A lay reader seeing "epiphenomenon" beside a drug will read it as "the drug does nothing". The caveat exists only in the Final score section. Label retained; qualifier added at first appearance. M/S/D and role do not alter the level. |
| A8 | Minor | Falsification and pitfalls, "Strict microdose equivalence" paragraph | "No experimental falsification of exact-dose efficacy was examined." -> "No experimental falsification of exact-dose efficacy was examined. Randomised dose-response evidence nevertheless runs against comparability: weight loss rose monotonically from 0.05 to 0.4 mg daily, weekly dose-finding was dose dependent, and 2.0 mg exceeded 1.0 mg weekly, while gastrointestinal events were dose related. Lower doses plausibly reduce adverse effects but, on every randomised dose comparison examined, also reduce weight loss." | PMID 30122305 (dose-ordered -6.0% to -13.8%; GI events dose related); PMID 26358288 (dose-dependent weight; GI AEs rose with dose); PMID 34293304 (SUSTAIN FORTE: -6.9 vs -6.0 kg, 2.0 vs 1.0 mg, trial-product estimand). All already cited. SC3 stays Α, but the text implies the equivalence question is merely untested when the in-review dose-response evidence leans against it. |
| A9 | Minor | Cap audit: five "field-valid definition" blocks, thirty "literal Γ hard rule" paragraphs, and the five Δ ladder lines | Verbatim aging-rubric text (lifespan/healthspan extension; ROS, NAD, mitochondrial and senescence hard rules) printed thirty-five times as the operative criteria, each answered "Satisfied: no" or "Not triggered" -> one sentence stating that rubric 06 (metabolism-aging) is used by adaptation, that its lifespan criteria and aging-marker hard rules do not apply to a weight-loss claim, and that the applied adaptations listed for SC1-SC5 are the operative thresholds. | Score card `rubric_routing.primary_rubric` = "06-metabolism-aging.md"; the review itself says "Lifespan is not imposed". Printing "Requires at least one intervention changing lifespan or healthspan ... Satisfied: no" for a weight-loss claim reads as template bleed and is misleading (SELECT [39] did change a hard healthspan outcome at 2.4 mg). Declared adaptation, so not Major; no level change. |
| A10 | Editorial | Stability; Knowledge gaps | Add: "Named de-escalation trials are registered (REST, NCT07294950, gradual reduction vs cessation, primary completion 2028-09; NCT07783854, tapering vs maintenance non-inferiority, primary completion 2027-12). They test tapering from 1-1.7 mg, not sustained sub-0.5 mg maintenance versus placebo, so Stability remains Stable." | ClinicalTrials.gov records retrieved 2026-11-02. They bear on the review's own proposed de-escalation arm. Not a change of value. |
| A11 | Editorial | PROPOSED DECISIVE EXPERIMENTS, Readout/Controls/Supporting/Falsifying line | "outcomes.. Controls", "measurement.. Supporting", "SC3.. Falsifying", "equivalence.." -> single full stops | Renderer joins pre-terminated strings with an extra period. |

**Rejected findings (checked, no change).**
(i) Best-supported rival label: `competing_models.models_evaluated[0]` is M1 (`is_nominal: true`, rank 3); the AT A GLANCE names M2 (rank 1, non-nominal, Ε). GM-1 applied correctly.
(ii) Aggregation: primary-subclaim on SC2 = Γ matches the headline; weakest-link sensitivity Α = min(Γ, Γ, Α, Γ, Γ) is correctly printed.
(iii) Β/Γ/Δ boundary for SC2: no examined or newly found source assigns sustained weekly sub-0.5 mg maintenance against placebo for 24-68 weeks; A1's EHR cohort is observational. Γ stands; Β was considered (A1 and A8 lean against the claim) and rejected because short weekly dose-finding, the exposure-response model and exposure-matched daily data give partial system support beyond co-occurrence.
(iv) Tournament fairness: M3 (behavioural cointervention) and M4 (response-selected dosing) are genuine alternative explanations of apparent low-dose success, not strawmen. M2 at Ε is compatible with the nominal claim rather than a strict rival; acceptable as the framework ranks explanatory models.
(v) All checked figures (STEP 1 1,961 / -14.9% vs -2.4% / 86.4%; STEP 2 -9.6% vs -3.4%; STEP 3 -16.0% vs -5.7%; STEP 4 803 randomised; FLEX-SEMA AE 56.8% / 32.6%, discontinuation 6.3% to 10.5%; INFORM 550; SEMALEAN 115/106, +4.5 kg grip; Ren 220 vs 212; Tzoulis 16 and 9 maintained at 1 and 2 mg; Wong 30 adults, every other week at existing dose; Bolek 20 patients, 0.25-0.5 mg, five months; STEP UP 55 scanned, 6 placebo) match their abstracts.
(vi) GM-2, GM-3 and GM-4 confirmed present in the review and score card; no residual fused numeral tokens outside the quoted GM-4 table.
(vii) Momentum "Associational only" and Stability "Stable": correct values; see A1 and A10 for content additions only.
(viii) No retraction, expression of concern or collective-author defect.

**Scorecard change.** None applied. Proposed only: add PMID 42857675 and 37424165 to `evidence_inventory`/`citation_ledger` (A1, A2); `reader-summary.json` limitations (A5).

**Residual limitations.** Spot checks were at abstract depth except where the prior pass had body receipts; FLEX-SEMA body figures were not re-read. The EHR cohort in A1 was read at abstract depth only.

**Approval:** Not approved as shipped pending A1; approval recommended once A1 is applied and A2-A9 are applied or explicitly declined by the lead. Level Γ and confidence Moderate are endorsed. - Claude Opus 5.5 (Anthropic), 2026-11-02.