Polybot

Pipeline counts + live-grid coverage. Covered = bot's LM version is registered in lm_features_registry (decay audit possible). Uncovered = run /blind-spot-hunt to register.

Beowulf v4 hypothesis registry. Lifecycle: discovered ▶ replay ▶ forward ▶ codify. Discovery via /blind-spot-hunt skill. G1 replay-gate runs LOCALLY (./tools/local_ops.sh blind-spot-gate <id>). Forward-validation watcher hourly checks 7d-old forward_pending rows.

Forward-passed blind spots awaiting G2 manual review. For each: confirm the feature is measurable in production, has no leakage, is marginally informative beyond existing LM features, is stable, and is backfillable. Then Codify (mark for inclusion in next LM retrain) or Reject as not-economic.

Each row is a deployed LM model. Click a version to see its features and Beowulf-flagged blind spots. Sync pulls feature list + coefficients from the live .pkl artefact and stamps the registry; rerun after any retrain.

Per-feature lifecycle: active (in model), decay_review (Phase 9 ablation flagged it as dead weight), codify_pending (will be added in next retrain), retired (was active, removed). Click a feature row to see its blind-spot history.