All exit/loss experiments (gate exemption, drawdown penalty, stop-loss,
asymmetric shaping) caused either exit spam (flat_l>35%) or training
collapse. The gate blocking FlatL was a FEATURE: it forces the model
to hold positions and learn from them.
Reverted to the proven dd049d9a4 configuration (wr=0.567) with only
the VALIDATED improvements kept:
- Reward chain wiring (apply_reward_scale)
- Adaptive C51 atoms [-3, +1]
- Adaptive LOSS clamp (observed ratio)
- Perf fixes (sync removal, memset, bg writer, TF32)
- Checkpointing
The loss minimization problem (L/W=1.32) needs unrealized PnL tracking
in the lobsim, not reward shaping hacks. Tracked for future work.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
f7427b27de