# Adversarial review log - protein-muscle-aging (shipped review)

Canonical file: `protein-muscle-aging.review.md` (batch `hsef_v53_traffic_first_five_2026-10-12`).
This log records an independent round-3 pass. The review already carries an embedded `## ADVERSARIAL REVIEW AND APPROVAL`
section from a round-2 pass dated 2026-10-14 (findings B-1 to B-6, all applied). Nothing in that section was taken on trust:
the reference list, the load-bearing claims and the scoring were re-derived against PubMed and the score card. Corrections
below are **proposed only**; no review, score card, page or flag file was edited in this pass.

Level before -> after: Ε -> Ε (unchanged). Confidence Moderate -> Moderate (unchanged; one in-scope primary-endpoint trial is missing from the inventory, which changes the boundary map and the reader summary but not the letter or the band).
Findings: 1 Major, 5 Minor, 1 Editorial. PMIDs verified 56/56. Claims spot-checked 18. Validator: not re-run (no file changed).

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1 (as stated in the review's own header; the package's `review_metadata.reviewer` and
`provenance_header.generated_by` fields record "Codex; HSEF v5.3"). **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-11-02.

**Scope of checks.** All 56 REFERENCES PMIDs verified against PubMed `get_article_metadata`: 56/56 resolve to the stated
first author, title, journal and DOI; 0 failures; no article-type field carries a retraction, retracted-publication or
expression-of-concern flag. Year offsets of one (e-pub versus issue year, for example references 7, 9, 10, 18, 24, 28, 31, 32,
35, 38, 48, 55) are the accepted convention and not errors. Eighteen load-bearing claims were spot-checked against the cited
abstracts, and against PMC full text where it exists: Park's 0.8/1.2/1.5 g/kg dose trial (n=120, 70-85 y, 12 wk, ASM/SMI primary,
gait speed secondary 0.09 vs 0.04 m/s, P=0.039, 1.2 g arm null); Tieland nutrition-only (n=65, SPPB interaction P=.02, no mass
change, leg-extension P=.059); Tieland training (n=62, NCT01110369, lean mass interaction P=.006, no strength or SPPB
interaction); OPTIMen (n=92 men, usual intake at or below 0.83 g/kg/d, controlled diets, LBM primary, strength, power, walking
speed and stair-climb null, fat mass the only difference); CALM (n=208, whey and collagen alone null versus carbohydrate, heavy
training with whey improving qCSA and strength, no training-only arm); PROVIDE (n=380, handgrip and SPPB primaries null, chair
stand and appendicular mass secondaries positive); Björkman (n=218, 12 months, no performance, grip or muscle-index benefit, 56%
mild GI effects, no post-trial mortality difference); Verreijen 2015 (n=80, intake 1.11 vs 0.85 g/kg, appendicular mass
preserved, function without between-group difference); Verreijen 2017 (PMC full text: n=100, ages 55-80, intake 1.13 vs 0.98
g/kg, no FFM effect, protein-by-exercise interactions for handgrip P=0.030 and 4-m gait P=0.045); bed-rest leucine (leg lean mass
partly preserved, function not); PRO-Elderly (n=200, no whey, soy, leucine or creatine add-on to training); Leenders 2013 (no
15 g/d add-on in healthy trained elderly); Katsanos 2006 and Pinckaers 2024 tracer composition effects; collagen 2015 (silica
placebo; positive FFM and isokinetic strength); Kang 2019 (abstract self-describes as case-control); Bagheri 2026 (women aged
40-77, protein-target null); Lim 2026 (n=45, 10 weeks, eSPPB and grip positive). **Every checked statement matched its source
in direction, sample size, endpoint status and effect.** Scoring logic was re-derived from the score card: SC1 is capped at Γ
by the biomarker-only rule before aggregation; the preregistered primary-subclaim rule selects SC3 = Ε; weakest-link
min(Γ, Ε, Ε, Γ) = Γ and scope-matched Ε are correctly reported as sensitivities; the Β/Γ/Δ boundary check carries
PMID 30475969 as the qualifying perturbation for SC2 and SC3; `causal_tally_audit` (1 positive unit, 23 screened
discriminating contrasts) matches the inventory, where `decisive_perturbation: true` holds for exactly the eight printed records
([1], [2], [3], [14], [22], [28], [53], [54]) and `discriminating_contrast: true` for 23; the role derivation (step 5,
reinforcing/stabilizing) and the M/S/D values (Mechanistic only / Stable / Contested) modify neither level nor confidence.
A targeted PubMed search for contradictory or omitted trials was run (PROMISS, pooled malnutrition IPD, recent network
meta-analyses).

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| P1 | Major | Evidence inventory and every boundary statement that depends on it: AT A GLANCE headline, "Belief versus evidence", COMPETING MECHANISTIC MODELS Rank 2, SC3 matrix row, REVIEWER NOTES, `reader-summary.json` overview | The PROMISS trial (PMID 34609621) is absent from the ledger, the inventory and the prose. Proposed: add it as a ledger row and a discriminating contrast (direction `supports`, not decisive-perturbation, caveats below), and amend the headline "Higher protein intake can improve selected muscle outcomes in some nutritionally vulnerable older adults; adequate-intake cohorts often gain little, and lean-mass changes do not establish stronger or more functional muscle." to "Higher protein intake can improve selected muscle outcomes in some older adults - most clearly in nutritionally vulnerable groups, with one open-label trial also reporting a walking benefit in community-dwelling adults whose habitual intake was near the allowance; other controlled trials at similar intake find no benefit, and lean-mass changes do not establish stronger or more functional muscle." Amend the REVIEWER NOTES sentence "The direct positive gait result is secondary to a mass-powered trial and needs confirmation with a functional primary endpoint." to "The Park gait result is secondary to a mass-powered trial; PROMISS (PMID 34609621) supplies a functional-primary positive in a different, non-malnourished population, but it is open-label, co-increased energy intake and lost significance per protocol, so independent blinded replication is still needed." | Retrieved and read in PMC full text (PMC8490609): 276 community-dwelling adults aged 65 or over, habitual intake below 1.0 g/kg adjusted body weight/day (mean baseline about 0.8 g/kg per the companion reports PMIDs 36678203 and 37960220), randomized to dietary advice plus protein-enriched foods targeting at least 1.2 g/kg/d versus no advice for 6 months. Primary outcome 400-m walk time improved by -12.4 s (95% CI -21.8 to -2.9) in PROT; leg-extension strength +32.6 N (10.6 to 54.5). Caveats the ledger row must carry: unblinded; control received no advice; energy intake rose 107 kcal/d; only 37.5% reached the target and the per-protocol walk-time effect was not significant; the PROT+TIMING arm was -4.9 s (CI crossing zero); secondaries are exploratory. Confirmed as the trial's main result by the secondary report PMID 37347495. This is an in-scope (age, above-RDA contrast, 6-month horizon, functional primary endpoint) randomized quantity contrast, so it is exactly the "functional primary endpoint" confirmation the review says is missing, and it falls in the population the headline describes as "often gain[ing] little". It also cuts against the review's baseline-intake boundary: OPTIMen's participants likewise had usual intake at or below 0.83 g/kg/d and were null. **Level unchanged:** with PROMISS the SC3 Ε crossing gains a second, context-distinct positive, but the effect is open-label, energy-confounded and unreplicated, and OPTIMen and the long supplementation trials remain null, so "reproducible intervention effects" for Ζ is not met. **Confidence unchanged:** adding a positive that conflicts with a null at similar baseline intake leaves reproducibility where Moderate places it. |
| P2 | Minor | "Upward threshold ladder", SC2 Δ, SC2 Ε, SC3 Δ and SC3 Ε lines | Each reads `Satisfied: yes.` and stops. Corrected to append the supporting record the score card already lists, e.g. for SC3 Ε: "Satisfied: yes. [22] 0.8/1.2/1.5 g/kg trial in 120 malnourished pre-frail or frail adults: 1.5 g/kg improved gait speed versus 0.8 g/kg (0.09 vs 0.04 m/s, P = 0.039), a secondary endpoint in a trial powered on appendicular muscle mass; the 1.2 g/kg arm was null." | `thresholds_tested` in the score card carries `supporting_evidence: [22]` for all four crossings; the prose drops it, so the headline-setting threshold is printed with no evidence at all. Spelling out that the only cited Ε crossing for SC3 is a single secondary endpoint at P = 0.039 is also what makes the Moderate band intelligible. |
| P3 | Minor | COMPETING MECHANISTIC MODELS, Rank 1 | "CALM, PRO-Elderly and healthy 24-week training studies show exercise benefits with null protein add-ons" to "PRO-Elderly and a healthy 24-week training study show exercise benefits with null protein add-ons, and CALM shows null protein-only arms alongside effective heavy loading" | CALM (PMID 33564844) has no training-only arm, so it cannot show a null protein add-on to training; the same paragraph's last sentence, the AT A GLANCE rival text and the Evidence section all say so. The sentence contradicts them. |
| P4 | Minor | `reader-summary.json`, `limitations` | "56 primary reports with 4 selected partial-body examinations; density, heterogeneous controls and incomplete access set the confidence, not correctness." to "56 primary reports with 4 selected partial-body examinations. Confidence is Moderate because the trials differ in formulation, achieved intake, comparator and endpoint, and positive and null results do not reproduce across baseline contexts; access depth is recorded separately and does not set the band." | Round-2 finding B-3 struck access depth from `final_score.confidence_rationale` under the lead's confidence-band policy, and the score card now carries the corrected rationale. The reader-facing summary still prints the struck wording, so the public page and the score card disagree on why the band is Moderate. The same file also carries `"redteamReviewed": false` although the review is marked approved after adversarial review; whether that flag is meant to track this process was not established here, so it is reported for the lead, not corrected. |
| P5 | Minor | COMPETING MECHANISTIC MODELS, Rank 2 | Append: "A pooled individual-participant analysis of nine RCTs in malnourished or at-risk older adults (PMID 37432139) found no overall effect of nutritional interventions on handgrip strength, and OPTIMen and PROMISS - with similar near-allowance habitual intake - point in opposite directions, so baseline nutritional status is not yet a validated responder rule." | Tournament fairness. Rank 2 is presented with its supporting positives and only a generic caveat; the pooled IPD null (van Zwienen-Pot 2023, Nutrients, DOI 10.3390/nu15092025, read at abstract level: OR 1.11, 95% CI 0.78-1.59 for a 3 kg or greater handgrip gain) is the most direct contrary evidence for this model and is not cited. Its interventions were energy-and-protein packages, so it bounds the model rather than refuting it. Rank order need not change. |
| P6 | Minor | Subclaim matrix SC3 row (the row that sets the headline), and by extension the remaining fused-token passages | "A quantified0.8/1.2/1.5g/kg trial in baseline malnourishedpre/frail elders shows a modest gait benefit at1.5g/kg over12weeks, with the intermediate armnull. A frail protein-only24week trial also reportsSPPB benefit while mass is unchanged, though its exact totaldose is not verified in the original abstract. Controlled0.8vs1.3g/kg feeding in stable function-limited men and one-year/sarcopenia trials are null." to "A quantified 0.8/1.2/1.5 g/kg trial in baseline malnourished pre-frail or frail elders shows a modest gait benefit at 1.5 g/kg over 12 weeks, with the intermediate arm null. A frail protein-only 24-week trial also reports SPPB benefit while mass is unchanged, though its exact total dose is not verified in the original abstract. Controlled 0.8 vs 1.3 g/kg feeding in stable function-limited men and one-year sarcopenia trials are null." | Carried forward from round-2 B-1, which repaired prose sections but left about 480 fused tokens in the HSEF apparatus, inventory and evidence cards pending an upstream regeneration that has not happened. The SC3 row is the headline-setting cell of the main matrix and is read by lay visitors; it should not wait for a batch regeneration. No content added or removed. |
| P7 | Editorial | "Cap audit" field-valid definitions (SC1-SC4) and "Upward threshold ladder" Δ lines | Prefix each literal quotation with "Literal shipped metabolism/aging rubric text (lifespan/healthspan wording); the applied clinical adaptation is stated below:" | The four subclaims each print "Genetic or pharmacologic intervention producing reproducible lifespan/healthspan extension ... Must separate developmental effects from true aging effects", and SC3 Δ is then marked satisfied by a 12-week gait change. The review declares the adaptation in "Rubric fit" and the applied-criterion lines, so this is not template bleed, but a lay reader meeting the ladder first will read it as a claim that protein extends lifespan. Label only; no scoring change. |

**Rejected findings (checked, no change).**
(i) Aggregation: primary-subclaim selects SC3 = Ε; weakest-link Γ and scope-matched Ε are correct sensitivities; caps precede aggregation (SC1 biomarker-only Γ). Headline Ε follows.
(ii) Tally: 1 positive qualifying unit, 23 screened discriminating contrasts and 8 printed decisive perturbations are three different counts and all match the score card inventory. With P1 applied the discriminating-contrast count becomes 24; whether PROMISS also earns positive-unit credit depends on how the ledger treats open-label advice-plus-food interventions, which the review's own rules leave to the author.
(iii) Verreijen 2017 secondary grip and 4-m gait interactions: abstract-silent, but confirmed in PMC full text (P = 0.030 and P = 0.045). The round-2 residual limitation on this point is closed.
(iv) Template bleed: the only off-topic term is "HIV", which appears once inside rejected finding R-10, where it is the subject of the check. No unrelated-hypothesis text found elsewhere.
(v) Momentum "Mechanistic only" and Stability "Stable": the B-5 and B-6 lead adjudications are applied consistently in the AT A GLANCE table, the M/S/D section and the score card. PROMISS is completed work and does not change either value.
(vi) Kang 2019 (reference 29) is indexed by PubMed as an RCT while its abstract calls it a case-control study; the review flags this and does not count it as decisive. Correct.
(vii) Bagheri 2026 (reference 17) enrolled women aged 40-77, largely outside the 60-or-over scope; it is not used as a decisive or positive unit, so no correction is needed.
(viii) Collagen 2015 is correctly held out of the clinical crossing pending the corrigendum contents.
(ix) Role of mechanism (step 5, reinforcing/stabilizing) is derived from causal direction, not tournament rank, and does not alter the level.
(x) The CALM registry number (NCT02034760) is not printed in the review, so round-2 B-6's statement about "all five trial registry identifiers in the document" remains accurate.

**Scorecard change.** None applied. P1 and P4 would require ledger and `reader-summary.json` edits if accepted.

**Residual limitations.** Recent network meta-analyses retrieved in the search (PMIDs 42942938, 42712362) pool mostly
multi-ingredient or exercise-combined formulations and do not isolate the above-RDA quantity contrast, so they were not used
as findings. The Ten Haaf 2018 trial that PROMISS cites as a third low-intake trial was not separately retrieved and is not
asserted here. Park 2018 and Tieland 2012 have no PMC full text; their figures were checked at abstract depth only.

**Approval:** Approval of the review is conditional on P1. The level and confidence stand, but the published boundary
statement ("adequate-intake cohorts often gain little") omits an in-scope primary-endpoint positive and should be revised before
the next edition. - Claude Opus 5.5 (Anthropic), 2026-11-02.