[zero-aska]

Closes #59.

What this is

One protocol pull request: a pre-registered exploratory analysis plan for the frozen-cohort candidate subset with no observed discovery record (13 sensitive / 8 refractory), written before any real-data Pilot 001 number exists. It freezes endpoints, permitted inputs, decision rules, uncertainty methods and stop rules, says which claims this n can and cannot support, and labels every claim.

• protocols/pilot-001-exploratory-subset-plan.md — the plan, on the research-question template's fields, with a source-fixed analysis contract in the style of the score-verification specification.
• tools/plan_exploratory_subset_design.py, registries/pilot-001-exploratory-subset-design.json, tests/test_exploratory_subset_design.py — the plan's design values, computed offline from the documented counts so that nobody invents a statistic after seeing numbers: the exact Mann–Whitney null for 8 vs 13 and 11 vs 17, its one-sided critical values, exact Clopper–Pearson intervals for every attainable sensitivity and specificity, and the frozen bootstrap/threshold parameters.
• CLAIMS.md — V28 (design values), K30 (published frozen-independent facts), K31 (subset counts), C22 (the hypothesis); counts line updated.
• Navigation and credit: README start-here row, roadmap paragraph, registries index entry, pointer in open-questions §4, CONTRIBUTORS note; verify.yml test count 310 → 317.

The contract in one paragraph

Evaluation subset: the 21 patients in the published 34-row Frozen_nonoverlap_pred_scores sheet whose patient key appears in no row of FFPE_Discovery_clinical. Primary estimand: empirical AUROC (half-credit ties, refractory positive). Primary decision rule: exact one-sided permutation test against 0.5 at α = 0.05 — with no ties, reject only when U ≥ 76 of 104 pairs (AUROC ≥ 19/26 = 0.7308; exact p = 0.04452, and 0.05210 one step below). Uncertainty: patient bootstrap (2,000 draws, seed 20261106, ≥1,000 valid draws) plus a DeLong interval. Threshold: fixed on the 40 other scored frozen patients at a 0.98 specificity target (0 allowed false positives on 29 negatives); achieved specificity and sensitivity on the 21 reported with the pre-tabulated exact intervals. The printed frozen values (AUC 0.79, 95% CI 0.63–0.96; sensitivity 0.25) are reported beside the result with no pass/fail. One sensitivity analysis on the 28 (17/11), no decision rule. Stop rules on hashes, row counts, joins, membership counts, score range, cross-check disagreement, enumeration totals and bootstrap validity.

Two tiers: Tier A evaluates the authors' released frozen scores (the pinned NLM workbooks of V08/V11, in memory, aggregates only). Tier B applies the same contract to the project's own reconstruction outputs — and the plan states plainly that nothing in the repository can produce those today (discovery-only reconstruction that keeps no fitted model; no frozen matrix pinned or mappable; no per-patient transfer route past the output allowlist), so Tier B needs a dated addendum and every open Gate 0 item.

What this cohort can and cannot support (from the plan)

Can: an exploratory above-chance test at the critical value above; descriptive estimates with honest width (every 95% sensitivity interval on 8 patients is ≥ 0.36 wide; a perfect 13/13 specificity gives a one-sided 95% lower bound of 79.42%). Cannot: specificity ≥ 98% by any bound (149/183 zero-error negatives needed, V09); confirmatory validation; reproduction of 0.83 or 0.79; population transportability; calibration; any clinical inference.

Claims cited, not replaced

V08 (released discovery AUROC 0.8088507878, rounds to 0.81, vs printed 0.83) and V23/C20 (discrepancy analysis) are cited as the governing records for the discovery question; no AUROC is invented for any subset.

Draft until the internal adversarial review of this change set completes; findings will be folded in before it is marked ready.

Checks actually run

• python3 tools/validate.py — PASS (registries, local Markdown links, attestation file untracked).
• python3 -m unittest discover -s tests -v — 317 tests OK (310 + 7 new).
• python3 tools/plan_exploratory_subset_design.py --check — artifact matches.
• python3 tools/check_implementation_hashes.py — OK (the new registry uses the dict form of implementation_sha256).
• Independent internal agent review of the design values and of the plan's factual claims against main (reader maps and an adversarial pass): in progress while this is a draft; findings are folded in before it is marked ready. This is internal review, not external scientific review.

What the checks cannot establish

No study file was read, no frozen score was seen, no threshold was selected and no AUROC was computed on any subset: the tests establish the design arithmetic and the artifact's provenance, not any model performance, patient independence (the 21 are outside the alignment fit but not untouched: P177 marker filtering, pooled reference channel, one institution) or clinical validity. The 13/8 class split is inherited from the 2026-10-25 bounded inspection by phrase pin, as the capacity record already does; the total 21 is replayable from the supplement audit.

Hard constraints honoured

No execution on study files; no Foundation-0/asma-data; no medical advice; no Pilot 001 run; no acquisition of study matrices; no contact or recruitment. Execution of Tier A needs the owner's explicit, dated instruction naming the plan version; Gate 0 Q5 alone is not read as that instruction. The parallel biology track (#57) is untouched; the two PRs both append to CLAIMS.md (different IDs: V27/K29 there, V28/K30/K31/C22 here) and will need a trivial merge of the counts line whichever lands second.

AI disclosure

Written by Claude (Anthropic) via Claude Code under human direction, with independent internal agent review. Internal review is not external scientific peer review.

---

Generated by Claude Code