vlabki/rr-speed-v4

Self-contained recurrent player-policy checkpoint.

• Source directory: best_reliable
• PPO update: 18250
• Environment steps: 112128000
• Action support: bc

Evaluation values below come from deterministic fixed-seed argmax runs. Frame averages exclude DNF races.

Best reliable evaluation

Metric	Value
Checkpoint update	18250
Cohort	20
Finish rate	95.0%
Finished mean frames	10706.58
Finished median frames	10697.00
Fastest finish frames	10645
Finished P90 frames	10726.60
First-place rate among finishes	100.0%
Mean wall events	0.65
Mean respawns	0.00

Best fastest evaluation

Metric	Value
Checkpoint update	17000
Cohort	20
Finish rate	75.0%
Finished mean frames	10687.47
Finished median frames	10681.00
Fastest finish frames	10633
Finished P90 frames	10712.60
First-place rate among finishes	100.0%
Mean wall events	1.30
Mean respawns	0.20

Files

Runtime weights, model config, normalization statistics, route reference, training config, and portable checksummed fallbacks are included. Raw rollout traces, optimizer state, and full training logs are excluded.