autocaliblisted

# EDPA Auto-Calibration — Monte Carlo signal-weight optimizer ## What this does Optimizes the five **signal weights** (`assignee`, `pr_author`, `commit_author`, `pr_reviewer`, `issue_comment`) in `plugin/edpa/templates/cw_heuristics.yaml.tmpl` against a synthetic corpus generated procedurally. The engine consumes those weights directly — there is no `role_weights` or `role_overrides` block any more (both were dropped in v1.11; see `plugin/edpa/scripts/engine.py:864`). The optimizer is self-contained: it generates its own ground truth via Monte Carlo, evaluates candidate weight vectors against it, and writes the best candidate back into the template when `--apply` is passed. ## Arguments `$ARGUMENTS` = optional flags forwarded to `calibrate_signals.py`. Common forms: - empty / `help` → show current calibration metadata, propose a default run - `quick` → adds `--quick` (200 MC samples; ~1 s; smoke test only) - a positive integer → `--scenarios <N>` (e.g. `2000`); default `1000` - `apply` → after calibration, write best weights back to the template - raw flags (`--scenarios 2000 --seed 7 --apply --report report.json`) → passed verbatim ## Argument resolution (when `$ARGUMENTS` is empty) 1. Read the current `calibration:` block from `plugin/edpa/templates/cw_heuristics.yaml.tmpl` and print: ``` Last calibration: method: MC random-sample + coordinate descent scenarios: 1000 records: 31041 baseline MAD: 0.0861 calibrated: 0.08