Context-compatible ML head replay

The earlier replay fitted candidate heads, but accepted provisional rubric changes dropped the active head. Only 9 of 87 pre-feedback predictions used a learned head. This was a lifecycle defect, not evidence that feature learning was ineffective.

The shared candidate fitter now rebuilds context-specific probability features, fits on trusted training labels, calibrates using out-of-fold predictions, and returns a complete context/head pair. Activation saves that pair together and rejects mismatched provenance. Failed feature collection leaves the incumbent unchanged. Below three training labels per class, fitting is explicitly deferred. Provisional acceptance is not proof of a measured improvement.

A fresh chronological replay used the same 87 existing votes and fixed split: 51 training, 18 development, and 18 protected audit items. Optimizer outputs were collected afresh and can differ from the earlier recording. The replay made 528 decision requests and 11 optimizer calls. The head was used continuously from Cycle 34 through Cycle 87: 54 predictions. Fitting occurred at Cycles 33, 40, 41, 60, 61, and 80. The final active model used nine independent probability features from five classifications. Each question's last class probability is recoverable from the remaining probabilities.

The endpoint audit made 36 additional decision requests and no optimizer calls. Both the final Jev-only path and learned-head path reused the exact same final context and cached answers. The protected audit contains 2 include and 16 exclude labels. It was not used for fitting or proposing context changes.

The accuracy and natural-weighted Brier gains do not establish improved alignment: the head missed both positives. The lifecycle repair is verified, but generalization of the learned decision boundary remains unresolved. Two positive audit items give very imprecise recall estimates. Do not tune against these protected outcomes or claim calibration quality from this comparison. The complete flywheel's prequential agreement was 77/87 (88.5%); that is a different statistic from final-model audit accuracy and does not isolate the head's contribution.

Prediction events now record the original decision-model probabilities, ML feature values, raw ML probabilities, calibrated final probabilities, and temperature. Confidence is the final probability of the selected class.

Private recordings are under var/head-lifecycle-replay-20261006-v1/: trace.json, runtime.sqlite3, run-comparison.json, and head-comparison.json. Article text and human explanations are not published in this note. The previous run is preserved. No new human labels were collected.