Roller PPO024 research checkpoint
Genuine native CPU physics / MPS PPO checkpoint 24, from our mps training run. Legal 61-observation input, 14-action output, rollers accessory, fixed command 1.1 m/s. No seed table, new privileged input or third-party opponent imitation was added.
Local original-evaluator fixed-command five-start compatibility score: 2.718 seconds, 5/5 finishes. Separate new ordinary-start validation: 128/128 finishes; command-stress validation: 64/64 finishes. All reported means are within the declared +0.030-second noninferiority budget against the general baseline.
This is a daily-run points candidate, not an improvement or podium claim. The published Roller best is 2.570 seconds and remains the better timed entry. A Roller run does not replace the Tug cup entrant. Official server scores can differ from local Mac results.
Model SHA256: f27c33e8b634f7b7bccf7dca93eae4d3e5ed116de4f156e7732ff16ef58c676e. Inherited upstream terms and attribution remain in upstream-base-card.md and LICENSE-NOTE.txt.