calibrate

You can review agent calibration drift by querying memory entries in JSON mode and applying human-gated approval to merge, reject, or defer behavioral updates into calibration files. Accepted changes are written to calibration files and an append-only lookback log.

Agent calibration drift consists of queued behavioral updates, rule overrides, threshold changes, and false positives or negatives emitted by subagents. You need drift management when subagents learn during sessions but lack direct write access to calibration files.

Update agent calibration files by presenting each drift entry alongside current calibration state, then choosing to merge, reject, or defer. The workflow maps memory keys to calibration sections and writes accepted changes to the file and lookback log.

Yes, the calibration review workflow deletes processed memory keys after merging and warns about stale scratch fallback artifacts. This cleanup prevents processed drift entries from accumulating and ensures scratch files are safely reconciled.

Yes, you can use this workflow to handle subagent rule overrides, threshold adjustments, false positives, false negatives, and dismissals. It maps these drift entries to calibration sections as either new rules or refinements.