Files
foxhunt/config
jgrusewski 55d70b7cc8 feat: dense micro-reward + DSR fix + counterfactual sign fix
Dense micro-reward: OFI momentum × price confirmation (MBP-10 mid-price
mark-to-market) × adaptive cost tolerance (capital_ratio × Sharpe_ema)
+ book aggression + retrospective hold quality bonus. Replaces flat
-0.0001 holding cost. Scale: micro_reward_scale=0.1.

DSR Sharpe EMA: was hardcoded price_change_dsr=0.0 — adaptive cost
tolerance was permanently floored at 0.1. Now uses actual per-bar
returns from ps[PREV_CLOSE_SLOT].

Counterfactual: cf_cycle==1 (magnitude) and cf_cycle==2 (order) now
undo do_flip before computing CF reward, then re-apply. Previously
2/3 of counterfactual experiences had wrong sign when do_flip=true.

Rank normalization threshold: 0.001 → 1e-5 for dense micro-rewards.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-20 00:03:51 +02:00
..