# Lab-task-delivery review — v1.3.0-rc.4

Mode: static benchmark; no training. Two tasks, three new responses per model per task. Acceptance uses clipped Spearman and valid complete rankings. This is a file/grader delivery, not a hosted-service SLA.

| Slot | Status and criterion | Measured evidence | Remaining scope |
|---|---|---|---|
| Q1B | Descriptive comparison supplied; default cross-model SD criterion PASS; formal corpus audit MISSING_EVIDENCE | Sample SD across Astra/Grok task means: abundance 0.338, dimerization 0.430. Criterion: SD >0.10 on a majority of the two tasks. Individual observations, means and sample SD are in results/summary.json. | Existing source/cutoff evidence retained; no claim of a complete private-corpus audit. |
| Q2 | PASS for supplied provenance and owner-confirmed use | 453 sequences and 906 labels, pinned source table and transformations; owner confirms intended use is licensed. | Rights evidence is owner attestation. Qualitative workflow/value rationale retained without invented economic scores. |
| Q3 | PASS for implemented grading contract, subject to accompanying replay receipt | Actual verifier, 44 controls, 28 indexed answers, complete worked example. Current native rewards agree with independent replay. | No additional scorer or top-k gate. |
| Q4B | PASS direct contract controls; M12/M13 audit MISSING_EVIDENCE | Full-permutation, reverse/shuffle/invalid and known-rho controls. | Formal protocol acceptance would require those audits or a separate adopted alternative. |
| Q5 | NOT_APPLICABLE | No model training or update. | None. |
| Q6 | PASS for observed EC2 execution and offline replay, subject to receipts | 12/12 current requests returned; 0 provider/infrastructure failures (0.00%). Explicit max/high, no tools, isolation receipts, concurrency 4. | Managed buyer deployment and current Modal integration are outside this file delivery. |

## Coverage and version

12/12 requests returned; ten rankings are valid. Grok omitted two IDs in one abundance answer and one ID in one dimerization answer; both receive zero. Another valid Grok dimerization answer has rho -0.001666, clipped to zero. All three runs remain in each model/task average. These are model-output failures, not infrastructure failures. See results/invalid-rankings.json. Six current model traces per task, three per model; all native configurations are in evidence/INDEX.json. The current campaign is separated from 23 historical native attempts. Historical model answers do not enter the new averages; only the frozen ESM baseline is reused. All fresh requests exactly match the v1.3.0 instructions and candidates.

The two endpoints share 453 variants. The ESM comparator uses 2,018 train and 412 validation variants with disjoint mutation positions. No training/both-mode partition manifest or forgetting test applies. No pass@k claim is made; rho >=0.80 is optional metadata rather than the accepted delivery gate.

## Remaining requirement

| Item | Status | Resolution |
|---|---|---|
| Generic M12/M13 ratios | MISSING_EVIDENCE | Needed only to claim full generic-protocol acceptance; direct task controls remain provided. |
| Corpus contamination audit | MISSING_EVIDENCE | Dated source evidence is supplied; private-corpus inclusion is unmeasured. |

## Evidence and reproduction

- [Mode and revision](MODE_DECLARATION.json), [acceptance](ACCEPTANCE.md), [provenance](PROVENANCE.md).
- [All native attempts/configurations](evidence/INDEX.json), [headline observations](results/scores.csv), [summary](results/summary.json), [replay index](results/regrades.json).
- [Commands and worked example](RUN.md); `python3 tools/validate_package.py` and `python3 tools/replay.py`.
- [Validation receipt](VALIDATION.json) records checks actually executed, including infrastructure smoke and current-model isolation.
