# Adversarial review log - sauna-cardiovascular-prevention (shipped review)

Canonical file: `sauna-cardiovascular-prevention.review.md` (package `hsef_v53_traffic_ranks_11_15_2026-10-13`).
This is an independent second adversarial pass. The review already carries an embedded approval dated 2026-10-14
(corrections A1-A3). That approval is not repeated or relied on here; every check below was re-derived against the
shipped review, its score card, `reader-summary.json`, `CONSENSUS_CROSSCHECK.md` and the published `reviews/` page.
No file other than this log and its JSON companion was written.

Level before -> after: Γ -> Γ (unchanged). Confidence Moderate -> Moderate (unchanged; rationale must stop counting out-of-scope reports, see S1).
Findings: 2 Major, 5 Minor, 2 Editorial. PMIDs verified 57/57. Claims spot-checked 21. Validator: not re-run in this pass; recorded package validator PASS (0 fail, 0 warn, 30 pass).

## ADVERSARIAL REVIEW AND APPROVAL

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

**Scope of checks.** All 57 REFERENCES PMIDs verified against PubMed `get_article_metadata`: 57/57 resolve to the stated
first author, journal and DOI; years agree within the print/e-pub ±1 convention (for example 28972808 is 2017 in PubMed and
2018 in print; 21173495 is 2010/2011); 0 failures; no retraction or expression-of-concern flag. Twenty-one load-bearing
claims were spot-checked against the cited abstracts, and one (the cold-shower trial, PMID 27631616) against PMC full text.
The checks concentrated on what holds SC2 at Γ: the KIHD event cohorts (25705824: 2315 men, 190 SCD, 407 fatal CVD, reference
group one session per week; 30486813: 1688 participants, 51.4% women; 28972808: 2277 men, gas-analysed fitness; 37248758:
2575 men), the Holter trial (15564698: 30 patients, 20 vs 10, PVCs, SDNN and BNP), WAON-CHF (27001189: 76 vs 73, BNP-ratio
primary not significant, NYHA/6MWD/CTR improved), the matched five-year Waon cohort (19304125: 64 vs 65, 31.3% vs 68.7%,
8 vs 12 deaths, matching without stated randomisation), the exercise-plus-sauna trial (35785965), the CAD null (37650138),
the CKD crossover (42113470), the HFpEF single-arm pilot and its planned controlled trial (41831305), plus 19154844,
31002283, 39004886, 35323077, 26174323, 40498100, 32192678, 38546822 and 41455014. Every checked number, direction, species
and system matched its source. Scoring logic was re-derived from the score card: caps precede aggregation; the
preregistered `primary-subclaim` rule returns SC2 = Γ, matching the headline; weakest-link (SC3 = Β) and scope-matched (Γ)
sensitivities are correct; the decisive-perturbation count (16) matches the 16 inventory records flagged; M/S/D do not
touch the level. A targeted search for omitted literature was run (PubMed).

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| S1 | Major | Subclaim matrix SC4 and SC5; Competing models Rank 2; Rubric fit (rubric07 routing); AT A GLANCE "Best-supported rival"; References 31-56; score card `final_score.confidence_rationale` | Two subclaims, one rival model, the secondary rubric and 26 of 57 references concern cold-water immersion, cold showers and vagus-nerve stimulation devices, none of which is in the registered claim. Corrections: (a) AT A GLANCE "Randomized exercise-plus-sauna35785965 demonstrates that heat can add measured effects beyond exercise allocation, whereas sham-controlled TREAT-AF32192678 demonstrates a different exact electrical intervention on objective AF burden. These distinguish real scoped effects from a universal interchangeable wellness narrative; neither identifies sauna mortality prevention." to "Randomized exercise-plus-sauna35785965 shows that heat can add measured blood-pressure and estimated-fitness effects beyond exercise allocation, and the CKD crossover42113470 shows acute vascular and walking gains. Neither tests hard cardiovascular events, so they identify scoped physiological benefit, not sauna event prevention."; (b) SC4, SC5 and Rank 2 either moved to a separately registered claim or each prefixed "Out-of-scope companion; not evidence for or against the registered sauna claim."; (c) confidence rationale "Multiple independent studies across 57 primary reports" to "Multiple studies across the 31 heat/sauna reports (refs 1-30 and 57), of which the hard-event evidence is one Finnish cohort family and one matched Waon cohort" | `tools/hypothesis_register.json` registers the claim as "Regular sauna use causally reduces cardiovascular clinical events in adults." and the score card `claim_restatement.split_into` is empty. Score card inventory: 16 records relate only to SC5 and 10 only to SC4. The HSEF competing-model rule requires each alternative to "explain the same core phenomenon the nominal hypothesis is trying to explain"; "Cold and vagal interventions have separate mechanisms and indications" explains no sauna-event observation. The lay-facing "Best-supported rival" box cites an AF-burden device trial (32192678: tragus stimulation, 26 vs 27) as context for sauna. Headline level unaffected (primary-subclaim SC2; weakest-link Β is set by SC3), but the confidence rationale and the lay framing are inflated by sources outside scope. |
| S2 | Major | AT A GLANCE table "Causal role"; Final score paragraph; score card `role_of_mechanism`; published `reviews/sauna-cardiovascular-prevention.html` | "epiphenomenon" (rendered to readers as "This travels alongside the process without causing it or being caused by it.") to a non-assertive classification. Proposed text for the role cell and lay gloss: "Not determined for hard events: sauna causally changes blood pressure and some heart-failure function in randomized trials, but whether it reduces heart attacks, strokes or cardiovascular deaths has not been tested causally." Because the frozen five-value vocabulary has no undetermined value, the exact category needs lead adjudication; "epiphenomenon" should not stand. | The review's own derivation says "no causal dependency of habitual sauna was directly established in either direction", which is an undetermined state, not "contributes no causal role". The review rejects Β because "controlled human thermal perturbation provides partial system support beyond co-occurrence", and the framework assigns epiphenomenon when the nominal is "supported only associationally". Randomized evidence shows sauna causally lowers systolic BP over exercise alone (35785965: -8.0 mmHg, 95% CI -14.6 to -1.4) and reduces PVCs in heart failure (15564698). `tools/build_v5_reviews.py` ROLE_PLAIN maps the value to the quoted gloss, which appears verbatim on the published page and tells lay readers the opposite of "unconfirmed prevention, not evidence that sauna is physiologically inert". |
| S3 | Minor | Subclaim matrix SC1; Competing models Rank 1 | SC1 "Sauna heat exposure changes cardiovascular physiology." graded Γ, while rival M3 "Controlled physiological and restricted functional benefit" is graded Ε. Correct by restating SC1 to what its rationale grades: "Sauna-induced physiological changes mediate a reduction in cardiovascular events." (Γ retained); or keep the statement and grade it Ε | The SC1 rationale says its Γ "concerns the prevention-mechanism inference" and that "the physiological phenomenon itself is not denied", so the printed statement and its grade disagree. The literal statement is met by randomized human interventions (35785965; 42113470 crossover, FMD 3.84 vs 6.68%; 39004886), which is context-specific causality (Ε) and is how the review itself grades M3. Headline unaffected (primary-subclaim SC2). |
| S4 | Minor | Cap audit "SC2 cap"; Adversarial steelman, Β-rejection sentence | (a) SC2 cap quote "Biomarker change without functional/phenotypic change caps at Γ." to the boundary-check ceiling ("No field-valid hard-event perturbation: ceiling Γ") plus universal cap 2, "Pure association evidence — however large, however replicated — cannot exceed Β.", stated as binding the KIHD event evidence. (b) "Β is rejected for the assembled mechanistic proposition because controlled human thermal perturbation provides partial system support beyond co-occurrence." to "Β is rejected for SC2 because its own scope contains randomized human thermal interventions with disease-relevant functional and rhythm endpoints in heart failure (WAON-CHF secondary NYHA and 6MWD; Holter PVC reduction) and a matched, non-randomized Waon cohort with fewer hard events; these exceed co-occurrence, but none is a field-valid hard-event perturbation, so Γ holds." | SC2 does contain functional/phenotypic change (27001189 NYHA/6MWD; 15564698 PVCs), so the quoted biomarker cap's trigger is not literally met. The steelman also concedes "If only the clinical endpoint were considered, a conservative reviewer could hold the nominal to association", yet the clinical endpoint is the preregistered primary subclaim. Re-derived on SC2's own evidence, Γ is defensible (Γ: partial system support, endogenous disease-relevant causal dependency not shown), so the level stands; the justification must rest on SC2, not on the assembled proposition. |
| S5 | Minor | Evidence and causal interpretation, KIHD paragraph; Knowledge gaps, first item | Add after "...without making the entire evidence base geographically independent.": "Incident (non-fatal) stroke shows the same inverse gradient in the KIHD men-and-women sample (HR 0.38 for 4-7 vs 1 session/week after activity and socioeconomic adjustment; PMID 29720543), which adds an endpoint but not an independent cohort. An independent Japanese cohort of 30,076 adults (JPHC) found near-daily hot tub bathing associated with lower incident CVD (HR 0.72, 0.62-0.84) and stroke, but not sudden cardiac death (PMID 32209614); it is a different heat modality and is observational, so it bears on geographic replication of the association, not on causation." Knowledge-gap item to read "...with repeated exposure and settings beyond the Finnish cohort family; the one independent non-Finnish cohort found concerns hot-water bathing, not sauna." | Both retrieved and verified in PubMed: Kunutsor et al., Neurology 2018;90:e1937-e1944, DOI 10.1212/WNL.0000000000005606 (indexed in PubMed with article type "Editorial" although the abstract is a prospective cohort); Ukai et al., Heart 2020;106:732-737, DOI 10.1136/heartjnl-2019-315752. Registered claim covers clinical events, not only deaths; every SC2 cohort source in the review is a fatal-event analysis. Neither source changes the level (association, capped at Β for the cohort evidence). |
| S6 | Minor | Cap audit, SC1-SC4 literal Γ hard rule "Single-species or single-strain evidence" (5 occurrences in the review, 6 in the score card) | "Any hamster mechanistic component remains model-scoped; human clinical trials and cohorts are evaluated separately." to "Not triggered: no animal-model source is in the inventory; all 57 records are human studies." | Template bleed. Score card `evidence_inventory`: `system` is "human interventional" for 52 records and "human observational" for 5; no hamster or other animal source exists in the review. |
| S7 | Minor | `reader-summary.json` "limitations" | "57 primary reports with 3 selected partial-body examinations; density, heterogeneous controls and incomplete access set the confidence, not correctness." to the score card's corrected wording, amended per S1: "Confidence reflects evidence density and reproducibility: multiple heat/sauna studies with partial replication but heterogeneous comparators, modalities and endpoints; reviewer access depth does not set the band." | The earlier pass corrected this exact sentence in the score card (A2) on the stated policy that access depth is not a confidence criterion, but the lay-facing reader summary still carries the uncorrected text. |
| S8 | Editorial | Score card `review_metadata.reviewer` | "Codex" to "Sol 6.1" (or a note reconciling the two) | The review banner and approval section both state "Generated by Sol 6.1" / "Authoring model: Sol 6.1"; the score card names a different author. Provenance should agree across artifacts. |
| S9 | Editorial | `CONSENSUS_CROSSCHECK.md` and `CONSENSUS_CROSSCHECK.html` (4 instances each) | "the2315-men", "absenceIRR0.71", "underlyingTable2", "OriginalTREAT-AF", "secondaryNCT04143269" to "the 2315-men", "absence IRR 0.71", "underlying Table 2", "Original TREAT-AF", "secondary NCT04143269" | Known numeral/acronym spacing defect. The earlier pass (A1, A3) fixed it in the review and video insert but not in these two public cross-check files. No value changed. |

**Rejected findings (checked, no change).**
(i) Tally "27 screened discriminating contrast units" against 28 inventory records with `discriminating_contrast: true`:
records [53] and [54] share study ID NCT04143269 (one trial, a primary and a secondary report), so 28 records are 27
study units. The number is correct.
(ii) Decisive perturbations: 16 printed ([7], [8], [9], [16], [17], [18], [20], [25], [28], [31], [32], [40], [47], [52],
[54], [57]) equal the 16 flagged in the score card.
(iii) Aggregation: preregistered primary-subclaim returns SC2 = Γ; weakest-link Β (SC3) and scope-matched Γ are correctly
reported as sensitivities. Β/Γ/Δ boundary: no field-valid hard-event perturbation, so Δ is correctly withheld.
(iv) Cold-shower Methods: the previous pass left "retrospective registration" and "unverified compliance" as unexamined. PMC
full text (PMC5025014) confirms both: registered with NTR5183 on 25 June 2015 "prior to data analysis" but after
recruitment, and "compliance to the intervention could not be verified". Resolved; the review is correct.
(v) Waon prognosis numbers and the "37.4 percentage-point difference, not 38% relative reduction" statement are correct
(relative reduction would be about 54%); "random allocation is not stated" is accurate to the abstract.
(vi) WAON-CHF: the abstract states the BNP change "did not reach statistical significance"; describing the primary
endpoint as null is accurate.
(vii) Momentum "Mechanistic only" and Stability "Stable": 41831305 states a controlled trial "is currently being planned";
no completed or underway work is mislabelled.
(viii) SC1-SC4 quote rubric06 lifespan/healthspan wording and SC5 quotes rubric07 psychiatry rules: this is a declared,
explicit rubric adaptation (Rubric fit paragraph), not undisclosed bleed. The substantive problem with SC4/SC5 is scope
(S1), not the quotations.
(ix) Reference years within ±1 (13 instances) and the subscript in reference 12's title: not errors.
(x) No collective-author or empty-DOI defect; every reference has a named first author and a resolving DOI.

**Scorecard change.** None applied. Proposed changes are confined to the findings above: `role_of_mechanism` (S2),
SC1 statement or level (S3), the SC2 cap text (S4), `final_score.confidence_rationale` (S1), `review_metadata.reviewer`
(S8). None alters the headline level or confidence.

**Residual limitations.** Spot checks rest mainly on abstracts; one paper was read in full text. The omitted-literature
search was targeted (non-Finnish cohorts, non-fatal events), not systematic. The S2 replacement category depends on a lead
decision about the five-value role vocabulary.

**Approval:** Approved conditional on S1 and S2 being resolved. The level (Γ) and confidence (Moderate) stand on
re-derivation; the two Major findings concern scope and a lay-facing causal label, not the grade. - Claude Opus 5.5 (Anthropic), 2026-11-02.