Complete gyro clock development evidence

complete-development.tar.gz contains 880 indexed files in 3,908,344 bytes. It retains both complete 402-module source capsules, all three prior/current problems, options, fits and native outcomes, current final normal models, schemas, logs, regression source and original producers. The nine-state, 63-factor comparison retains every residual and derivative block and performs zero optimizer steps.

The receipt and schema bind the original archive, index, source commits and native artifacts. The summary and schema retain complete input equality, every reported comparison, initial test/command/lint failures and remaining limitations. Baseline outcomes are reused from the previous stored batch cohort; both source capsules establish their binding to the baseline.

Seed	Current status / iterations	Complete native result
20261010	converged / 20	JSON
20261011	max_iterations / 32	JSON
20261012	converged / 19	JSON

All three quality policies pass and all 32 signed controls are detected per case. These are spent development outcomes, without a speed, independent accuracy or general convergence claim. Raw final gradients remain above the configured stopping threshold. The middle case stops early at 32 of its configured 60 iterations with max_iterations; the frozen solver also uses that status for damping exhaustion. This path is inferred from its control flow, without a captured damping trace.

The status erratum and schema supersede the original summary's misleading candidate_iteration_limit_cases key and the initial prose. The archive, summary, native outcomes and all their digests stay unchanged. The producer binds the saved fit/options/source and the initial failed CI job excerpts. Quality-policy passes and solver convergence remain separate; no reruns, retuned gates or numerical edits were used to resolve that test assumption.

See the full comparison, boundary convention and CLI replay. To replay an original producer, copy generator.py.txt from the archive to outputs/gyro-clock-development-v1/run.py at the numerical candidate, then use a fresh output directory and its declared source argument. The archive's factor producer similarly records its original working path and hashes. It does not need a GPU or an original recording.