# Adversarial review log — ldl-ascvd (revised edition)

Authoring model: **Sol 6.1** (this edition is the published synthesis of a cross-model red-team pass on 2026-09-30).
Adversarial reviewer for this round: **Claude Opus 5.5 (Anthropic)**. Date: **2026-10-14**.
Round 2 of the adversarial pass (brief: `deliverables/ADVERSARIAL_REVIEW_BRIEF_2026-10-06.md`, ROUND 2 section).

| Item | Value |
|---|---|
| Level before | Ζ — Multiple independent causal lines converge |
| Level after | Ζ — unchanged |
| Confidence | Moderate overall (High for SC2/SC3/SC5) — unchanged |
| Role | primary driver, derivation step 3 — unchanged |
| M / S / D | Mechanistic only / Stable / Mild overstatement (implied Η, assessed Ζ, +1) — unchanged |
| Subclaims | SC1 Η, SC2 Θ, SC3 Η, SC4 Ζ, SC5 Η — all unchanged |
| PMIDs checked | 39 / 39 resolve to the stated title, first author, journal and year |
| PMIDs failed | 0 |
| Retractions / EoC | none; 3 records carry errata, correctly attached to their parent rows |
| Load-bearing numbers spot-checked | 16, all matching digit for digit |
| Findings | 0 Critical, 1 Major, 0 Minor, 0 Editorial |
| Validator | PASS — 0 fail, 0 warn, 33 pass under the v5.4 validator this edition declares (30 pass under the archived v5.3 validator) |

## Findings

| ID | Severity | Location | Change |
|---|---|---|---|
| E1 | Major | "Consensus consistency validation — 9 October 2026" block, frozen-inventory line and conclusion; the same two sentences in `consensus-addendum.md` | "Frozen scorecard inventory: 33 evidence rows; 28 claims" → "39 citation-ledger sources, of which 33 are scored evidence rows; 28 claims", and "original sources/claims stay at 33/28" → "stay at 39/28 (33 of the 39 sources are scored evidence rows)". 33 is the length of the score card's `evidence_inventory`, but the sentences present it as the review's source count, contradicting the header, AT A GLANCE, `provenance_header.sources_total: 39`, the 39-row `citation_ledger` and the review's own "Fourteen of 39 records". Scoped rather than deleted so both inventories stay traceable. No grade, subclaim or claim count moves. |

**Revision-regression check.** The 9 October revision spliced in the Consensus consistency block, and finding E1 is **a regression introduced by that splice** — the earlier edition did not carry the 33/39 mismatch. Nothing else in the splice contradicts the preserved document, and it leaves the grade, scope, ledger and subclaims untouched.

## Systematic artifacts reported by the lead — all four checked

1. **False "earlier adversarial approval" boilerplate** — **present but true here, so not corrected.** Verified before changing: an earlier independent adversarial pass exists, the 2026-09-30 cross-model red-team pass recorded in this document's own synthesis note and at `reviews/ldl-ascvd/redteam/redteam-pass-2026-09-30/` (25 of 32 findings accepted; grade moved Η → Ζ, confidence High → Moderate).
2. **Confidence set from access depth** — **absent.** The rationale cites directionally inconsistent evidence from independent groups and heterogeneous assays (NMR LDL-P, immunoassay apoB, calculated non-HDL-C and remnant cholesterol), which is the framework's Moderate text almost verbatim; access depth is reported separately in the limitations section. The overall band tracks the binding subclaim SC4.
3. **"Best model" renderer bug** — **absent.** The score card's first model is the broader apoB-particle retention model with `rank_in_tournament: 1` and `is_nominal: false`; the nominal LDL-particle model is ranked 2, and AT A GLANCE records "Nominal-model rank: 2 of 5".
4. **Reviewer-notes / Consensus-splice contradiction** — **absent.** No assertion that Consensus work was not performed; the spliced block states its own limits.

Score card `review_metadata.reviewer` left as issued; corpus-wide house provenance.

## Rejected findings (checked, no change)

1. **Headline Ζ** — the registered sentence is conjunctive and the framework requires scoring the conjunction, not its easier half. Weakest-link across SC1–SC4 was pre-registered on 2026-09-29 and returns Ζ from SC4; the Η reading is reported, not substituted, and all three rules' results are stated. Η for SC4 is refused on rubric text plus a named contradiction from independent groups; Ε is refused because the alternatives are genuinely narrowed.
2. **Momentum "Mechanistic only" despite two active outcome trials** — the artifact names PREVAIL and ORION-4 with registry identifiers and enrolment and states why neither crosses SC4's boundary. Framework definition applied, not boilerplate.
3. **The Lp(a)HORIZON sponsor topline** — named for currency, flagged as not peer reviewed, excluded from scoring, and attributed to the broader apoB model rather than the graded wording.
4. **Collective-author reference defect** — references 4, 7, 8, 24 and 27 all carry group authorship and all render with the group named and no leading comma.
5. **Rendering defects** — none found.
6. **Errata on references 6, 22 and 23** — attached to their parent rows rather than given rows of their own, which is the required ledger discipline; none is the sole support for a score.

## Residual limitations

Access depth is the main constraint and the review states it accurately: 14 of 39 records read in complete full text, 25 at abstract or metadata depth, with no claim requiring text beyond the abstract attributed to the latter. Two items could not be re-verified here: the apoB change behind the evacetrapib null, which the review itself lists as an open knowledge gap and which is load-bearing for reading the CETP results as concordant; and the 2026-09-04 Lp(a)HORIZON topline, a sponsor announcement with no peer-reviewed report, correctly excluded from scoring. The rubric adaptation is the review's own largest caveat: no cardiovascular or lipidology rubric was supplied, so rubric 06's lifespan and healthspan language is read as hard ASCVD events and plaque trajectory.

## Files changed in this case folder

- `ldl-ascvd.review.md` (banner + ADVERSARIAL REVIEW AND APPROVAL section + E1)
- `consensus-addendum.md` (E1; sha256 `5937798e26978c3942ccff0b6ac6c0e5723fb35b20cc38f70de53cc4e158a9ef`)
- `validator-adversarial-2026-10-14.txt` (new)
- `SHA256.txt` (refreshed; keeps the package's two-entry shape, so the addendum hash is recorded here instead)
- `sol61-original/` holds the pre-review copies of the review, score card, consensus-addendum.md, SHA256.txt and validation-report.json.

`ldl-ascvd.scorecard.json` was **not** changed — no scoring field needed correction.
Package-level files (`README.md`, `COMPLETION.json`, `FINAL_VERIFICATION.json`) were **not** edited.

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