Complete batched screw evaluation evidence

complete-evidence.tar.gz retains the first complete twelve-worker comparison against 5278168adbe8d1cffc4abbcd13800d901d0139e7. The numerical candidate is df4144a4d135dea3b0863847f26bc778b56fc719.

The archive contains the baseline's 401-module and candidate's 402-module source capsules, frozen benchmark tools, derived public GLIM geometry, all twelve complete input/options/result/fit/timing records and logs, validation schemas, regression test sources and the original archive producer. The development record retains the initial smoke failure and corrected runs. The receipt and its schema bind the archive's exact size, digest, file count, source commits and index.

The summary and schema, together with the development record and schema, retain the original measured bytes and the comparison's limits. Paired median API ratios are 2.023x for spent synthetic joint recovery and 4.709x for spent public geometry. Including fresh-worker startup and output, they are 1.462x and 2.189x. One joint API pair is 1.641x; the near-threshold joint median does not establish a universal 2x improvement.

Replay without a GPU or original recording:

python tools/audit_screw_residual_archive.py docs/assets/screw_batched_evaluation/complete-evidence.tar.gz

The command exits 0 after validating the complete records and both 2x median API gates. The auditor checks safe paths, schema and digest bindings and full input/result equality without importing archived calibration code.

All paired inputs and complete results match exactly. Joint fits retain 21/23/29 iterations, policy passes and all 32 detected controls; geometry fits retain 15 iterations. Assigned GLIM times and capture-local nearest planes serve a processing fixture, without a physical trajectory or accuracy claim. See the full comparison and scope. This is CPU development evidence; it makes no independent accuracy, reliability or SOTA claim.