Flashback

Find the step a training run broke on, from 452 bytes per step.

The top chart is what you would see on your dashboard. The bottom chart is what a 452-byte-per-step sketch saw at the same steps. Everything below is computed in your browser from recorded telemetry of real training runs with real injected faults — the ground truth was written down before any detector looked at the data.

What the loss curve shows

What the sketch shows at the same steps

Earliest detection, by metric

Bisect by hand

Each probe asks the git bisect question — has anything gone wrong at or before this step? — and halves the bracket. That phrasing matters: “is this step bad?” is not monotone (a flipped bit is bad at exactly one step), so a binary search on it walks off the end of the run. “has it gone bad yet?” is monotone — and it is also the question you can afford to ask expensively, since in the deterministic replay mode each answer costs one state reconstruction.