BOLT-Safe Fleet Learning for Robot Cells

BOLT is a simulation of how factory robot cells could share fixes safely. In MuJoCo, Cell A drops a glass jar whose low friction is hidden from the robot. A monitor writes a failure report, and an open LLM on AMD Token Factory (DeepSeek-V4-Flash and Gemma-4-31B-IT) proposes a rule: a grip force and carry acceleration, or "no safe rule". The rule is not trusted. Check 1 replays it over 300 simulator seeds and passes only if the Wilson 95% upper bound on drop-or-crush is at most 0.40 with zero crushes. Check 2 is a human review. A default-deny permission gate then requires the source owner to allow sharing before another cell receives the rule. Every decision is written to a SHA-256 hash-chained audit log. On held-out seeds, failures fell from 100% to 18-30%, with zero crushes across all 10 rules (2,000 rule trials, 3 distinct policies). A plain 22.5 N heuristic had comparable task success but crushed the jar in 2/100 and 4/100 trials. The pre-registered 100-seed gate passed 0/10 rules; the 300-seed rerun passed 10/10 and is post hoc. Limits: one object class, simulation only, no hardware claims, and Check 2 was auto-approved in the evaluation run.