Add mixture-aware post-failure design diagnostic

[FlorianPfaff]

Scope

This is a separately configured post-failure diagnostic stacked on immutable PR16 artifact commit d1251dddeaa706d79feaaec99185dcc4236aa4a3. It does not alter the original stress failure or its thresholds.

The producer was fixed before evaluation at exact head 564ab3cfe42f67867d70a5d5869c7c4401d94c8e, with configuration SHA-256 338ac9eea51f37d09b46303bdec66d5acc8d5e6fb16740c9afb036de3063c5a7. The evaluation was executed once, without tuning or rerun, and frozen at exact artifact head 5dc8c9280c5d3e042ce25d926c0d856ed9776ad9.

Locked method

• chronologically earlier heuristic maximin N=60 allocation;
• three-fold cross-fitted held-out scoring;
• all 15 pairwise nonnegative-cone models;
• maximum-over-15 composite statistic calibrated separately under each matched pure candidate;
• candidate-specific cross-fitted residual lack-of-fit thresholds;
• null-calibrated maximum pure-over-null and winner/runner thresholds;
• exact upper conformal rank 191 from ceil((200+1)(1-0.05));
• independent audit Wilson-lower >=0.70 gates for pure-label retention, null rejection, each individual mixture-pair rejection, and fixed-probe rejection;
• disjoint calibration/audit/evaluation seeds 196613, 262147, and 324949, with 200 replicates each;
• immutable baseline/allocation provenance, clean Git enforcement, checksum-bound outputs, and an independent verifier.

An underpowered pure winner is forced to abstain. Pair/null/probe rates remain descriptive and are labeled mandatory_abstain_underpowered when that exact contrast fails its audit gate. Evaluation did not change thresholds, enablement criteria, or configuration.

Frozen result

Diagnostic component	Audit status	Evaluation result
Six matched pure candidates	6/6 enabled	correct-call rates 0.800–0.955; minimum Wilson lower 0.7391
Null	enabled	false-pure rate 0.005; Wilson upper 0.0278
50/50 mixture pairs	9/15 enabled	worst enabled false-pure rate 0.195; Wilson upper 0.2554
Remaining mixture pairs	6/15 mandatory abstain	descriptive false-pure rates 0.215–0.560
Fixed out-of-span probe	mandatory abstain	raw false-pure rate 0.730; Wilson upper 0.7868

The result is therefore partial. The targeted cone gate improves rejection for nine audited in-span pairs but leaves six pairs underpowered and fails the bounded out-of-span probe. It does not establish open-set robustness.

Provenance

Artifact path: results/design-mixture-diagnostic-n60/.

• SHA table: a3a9bccf6e9c597e360bab566ea7680053f3bb0c5dabd5aff2f6c9cc8744659f
• manifest: bf08f458a3f34339a62c8ae74fc0a1cafdf7f74f21e939d174df86b021bc21b7
• summary: 1311330a9f7b3278eb16dbb48d65b57b8337a695c6ad2026cac1efbff795b3c2

The independent verifier passed all nine payloads and bound the immutable baseline manifest.

Scope limit and verification

This remains targeted to positive two-candidate mixtures and one bounded nonlinear probe. It is not an arbitrary open-set, sequential-protocol, or biological robustness certificate, and it is not promoted as a main-paper robustness claim.

Producer 564ab3c passed pycompile, Ruff, mypy over 34 source files, 16 focused tests, and GitHub quality/build/Python 3.10–3.13. Artifact head 5dc8c928 passed the independent verifier, Ruff, mypy, and the same focused tests locally; its GitHub CI run 32810342063 is in progress.