jgrusewski
d980fdedf7
feat(generalization): #7 counterfactual + #20 lottery ticket + remove enable_ flags
Task 7 — Counterfactual Experience Augmentation:
Every timestep writes BOTH original AND mirror experience to output.
Mirror action: exposure_idx → (b0_size-1) - exposure_idx (L100↔S100).
Mirror reward: -reward. Output buffers are 2x for counterfactual.
Always active — no enable_ flag. Doubles effective data diversity.
Task 20 — Lottery Ticket Pruning:
lottery_ticket_mask kernel: zeros pruned weights in f32 + bf16 after Adam.
lottery_ticket_compute_mask kernel: magnitude threshold → binary mask.
At pruning epoch (50): read all weights, sort by |w|, bottom 70% → mask=0.
After that: mask applied every training step (one kernel launch).
Invalidates CUDA graphs on mask creation (new step structure).
Enable_ flag removal:
ALL enable_ conditionals removed from hot paths. One production path.
- enable_domain_randomization → removed (always randomize)
- enable_mirror_universe → removed (always mirror odd epochs)
- enable_vol_normalization → removed (always normalize)
- enable_anti_lr → removed (always anti-intuitive LR)
- enable_gradient_vaccine → removed (always project gradients)
- enable_causal_intervention → removed (always run interventions)
- enable_adversarial_self_play → removed (always saboteur active)
- enable_counterfactual → never added (always counterfactual)
Domain randomization GPU kernels: removed enable_jitter and enable_dr
parameters from CUDA kernel signatures. Kernels always randomize.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>