# Adversarial review log — trained-immunity

**Authoring model:** Sol 6.1. **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-10-14.
**Round:** 2 (2026-10-14). **Canonical path:** `consensus_consistency_audit_2026-10-09/revised/trained-immunity/trained-immunity.review.md`.

## Summary

| Level before | Level after | Confidence before | Confidence after | Critical | Major | Minor | Editorial | Approval |
|---|---|---|---|---|---|---|---|---|
| Ε — Context-specific causality established | Ε — Context-specific causality established (unchanged) | Moderate | Moderate (re-derived under lead policy; band stands) | 0 | 2 | 3 | 1 | Approved as revised |

PMIDs checked 63/63, all resolved to the stated title, first author, journal and DOI. Retraction / expression-of-concern
screen: none found. Load-bearing claims spot-checked: 13. Validator (`validate_review.py` v2.3, HSEF v5.3 tree):
**pass** — 0 FAIL, 1 WARN (`v5.claim_form` "phenomenon", rejected per standing convention), 31 PASS.

## Findings

| ID | Severity | Location | Original → corrected | Evidence |
|---|---|---|---|---|
| TI-1 | Major | Consensus provenance reconciliation block | "The earlier adversarial approval refers to its original edition" → "No earlier adversarial approval exists for any edition of this review" | No edition of this review carries an approval section or approval line. The 2026-10-06 adversarial pass covered 30 other slugs; this is one of the four cases in this group where the Oct 9 revision spliced the sentence into a review that had no such approval. |
| TI-2 | Minor | Evidence card [3] status line | "ErratumIn: PMIDnot supplied." → "Correction screen: no erratum, corrigendum or retraction record is indexed in PubMed for this report; status_checked_on 2026-10-14." | Renderer emitted the known "PMIDnot supplied" placeholder, and PubMed holds no correction record for PMID 25258083 (searched by author, journal, title and publication type). A status flag may not assert a correction it cannot identify. |
| TI-3 | Minor | Decisive-perturbation item [3]; card [3]; RESEARCH TIMELINE 2014 row | "linked author correction noted" (×2) removed; "linked correction" → "no indexed correction record" | Dependent prose repeated the unverifiable claim from TI-2 in three further places. |
| TI-4 | Minor | KNOWLEDGE GAPS, correction-resolution item | Grouped "de Laval, HIF-1α and microbiome/BCG" corrections → de Laval errata PMID 32386557 and PMID 36608675 and gut-lung Author Correction PMID 37161093 named; HIF-1α claim withdrawn | The three real corrections were verified against PubMed and attach to the right parents; the fourth does not exist. Naming identifiers makes the residual limitation checkable. |
| TI-5 | Major | AT A GLANCE (new **Confidence basis** paragraph); `final_score.confidence_rationale` and its two mirrored copies in the score card | "density, heterogeneous controls and **incomplete access** set the confidence" → rationale grounded in evidence density and reproducibility, with reviewer access depth explicitly excluded as a determinant | Lead policy on confidence bands, and `hsef-framework.md` § Confidence band. Step 2 of the lead's two-step test applied: access depth was **not** the only thing holding the band down — the review documents irradiated-recipient and conditioning confounds, few laboratories behind the purified-transfer result, induction-stage interventions standing in for maintenance tests, and a clinical arm split between replicated adult nulls and single-setting neonatal benefits. These are valid determinants, so **Moderate stands**. |
| TI-E1 | Editorial | RESEARCH TIMELINE, 2018 mevalonate row | "Constitututive" → "Constitutive" | Spelling. |

## Rejected findings (checked, no change)

1. **"Best-supported rival" renderer defect.** Score card shows `rank_in_tournament: 1` / `is_nominal: false` for the
   dose-and-context set-point model; the review's own nominal (M1) ranks 2. The defect did not occur here.
2. **Print year vs PubMed e-pub year ±1** — Domínguez-Andrés (Cell Metab 2019 vs 2018), Lechner (JACI 2022 vs 2021),
   Jensen (J Infect Dis 2015 vs 2014). Standing convention.
3. **Gaps at [4] and [30] in the decisive-perturbation inventory.** Numbered against the 63-row citation ledger; those
   two rows carry no discriminating contrast. 61 items, tally reports 61.
4. **Inline `PMID22901542` token form.** Systematic house token (used identically in the `Study:` ledger field across
   the corpus), not a dropped space inside a word.
5. **`v5.claim_form` "phenomenon" warning.** Standing convention: the registered claim states acquisition and
   persistence plus the enhanced-response consequence.
6. **Scope-matched aggregation misread.** Not the declared rule; and the headline is not raised above the
   scope-matched result, so the warned-about error cannot arise.
7. **REVIEWER NOTES vs spliced Consensus section.** No contradiction: the notes record two Consensus Pro queries,
   zero API/Deep, one SciSpace query and two denied Scite body attempts, consistent with the Consensus section.
8. **Momentum.** "Mechanistic only" is already the correct value; no "Decisive work underway" claim to correct.
9. **Collective authors.** No reference in this review has a consortium first author; no leading-comma defect present.

## Residual limitations

Access depth is unchanged: 8 partial-body and 55 abstract/status examinations, no complete-paper or supplement audit.
Three partial-body-sourced figures (Cirovic 15 vs 5 at day 90; Mills 18-week endpoint; preserved recipient lineage
proportions) could not be re-verified from abstracts and are not contradicted by any record examined.

## Changed files

| File | sha256 | bytes |
|---|---|---|
| trained-immunity.review.md | d626947975a3be372047863755175bc3fb182c0ec40ca69c5fdf736edc340f51 | 126836 |
| trained-immunity.scorecard.json | dd8a7ff4945f92e1f620837a3a951090091268133d8eeb7a84db5093c69efff4 | 361198 |

`SHA256.txt` regenerated; `validation-report.json` re-run. Pre-edit copies in `sol61-original/`.


## Collective-author screen (both known forms)

Screened for a dropped author leaving a stray leading comma, for a consortium name truncated to its last word so
that it reads as an ordinary surname ("Group et al.", "Network et al.", "Consortium et al."), and for dropped
nobiliary particles. Method: every reference's rendered first author compared against the PubMed record,
including records whose first author object comes back empty because the article is credited to a consortium.
Result: no collective first author in this list; particles intact ("de Laval", "Domínguez-Andrés", "O'Farrell", "Flores-Gomez", "Giamarellos-Bourboulis"). Five records carry a consortium **mid-list**, which does not affect the rendered "First-author et al." form. No defect.

## Line-ending note

The edit pass initially normalised these files from CRLF to LF. That was unintended and has been reversed: the
review and score card are back on CRLF, matching their pre-edit state and the rest of the package, and the hashes
below are the final CRLF files.

**Approval:** Following adversarial review and the corrections listed above, the revised document is approved. — Claude Opus 5.5 (Anthropic), 2026-10-14.