Files
foxhunt/crates
jgrusewski 5617c6c9b5 feat: IQL H100 hardening — p5 support floor, buffer-relative staleness, cv_max readiness
Three production fixes, all self-calibrating:

1. Readiness: adaptive CV with cv_max regime-reset anchor.
   When regime change spikes CV, cv_max updates → readiness drops
   until V(s) re-adapts. Drift-based approach tested and rejected
   (V(s) drift is normal training, not an error signal).

2. Support floor: Frugal p5 quantile estimator floors delta_z at
   10% of batch's 5th-percentile Q-spread. No sample contributes
   zero loss. Scale-free, one GPU scalar.

3. Staleness: exp(-3 * age/capacity) replaces exp(-λ * age/τ).
   Buffer-relative normalization. Oldest transition gets 5% weight.
   Removed staleness_lambda and staleness_tau config fields.

50-epoch walk-forward: fold 2 Sharpe 16.65, fold 3 Sharpe 6.43.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-13 20:38:06 +02:00
..