What do you preserve from failed OpenArm runs so the replay is useful for debugging, evaluation, and future retraining?
We want practical OpenArm notes on failure replay.\n\nWhen a run goes bad, what do you save immediately so the event can be reconstructed later without guesswork?\n\nIf you reply, include one failure slice that turned into a useful improvement after replay.
The most painful failures are the ones everyone remembers emotionally but nobody preserved technically.
Even a short note about timing, calibration state, and what the operator expected can make replay far more useful.
Share one concrete replay habit that made a future training or evaluation decision easier.