Incident Replay and Post-Event Reconstruction

Every shot and every upset is fully reconstructable: synchronized telemetry, twin state, model confidence, alerts, and operator actions replayed on one timeline.

Deterministic reconstruction

When something goes wrong — a breeder disruption, a burner plug excursion, an interlock trip — the team must be able to replay exactly what happened and what the machine, the models, and the operators each knew at every instant. Incident replay reconstructs a bounded window from the L2 petabyte telemetry archive, the twin's recorded state stream, the anomaly ensemble outputs, and the operator action log, aligned to a single PTP-disciplined clock.

What a replay contains

• Raw diagnostics at native rate (down to the sub-microsecond sampling)
• Twin state and its per-region confidence at each step
• Every alert: when it fired, its lead time, its recommended action
• Operator actions and acknowledgements, attributed and time-stamped
• Interlock and failsafe events from L1 with hardware timestamps

The replay is scrubbed on a timeline that drives the same 3D twin overlay used live, so an investigator watches the flux surfaces, wall temperatures, or plug density evolve exactly as they did — with the same confidence hatching that was shown at the time. Crucially, replay distinguishes what was knowable (in the data) from what was known (surfaced to the operator), which is how alerting and display gaps are found.

Counterfactual replay

Because the twin is a model, replay can branch: re-run the recorded inputs through an updated model or a different MPC policy to ask whether an event would have been caught earlier or avoided. This is a controlled offline exercise — it never touches the live machine — and it feeds the L0 retraining loop and MPC tuning. For the breeder, counterfactual replay is how disruption-precursor lead times are validated; for the burner, how plug-density controllers are compared.

replay = archive.window(shot_id, t0, t1)          # aligned multi-source pull
replay.attach(twin_state, confidence, alerts, ops)  # one PTP timeline
for t in replay.scrub(step=1e-3):                    # 1 ms scrub
    overlay.render(replay.twin_at(t), replay.conf_at(t))
# branch: what would a candidate model have flagged?
cf = replay.rerun(model=candidate_v)                 # offline only

Replays are the raw material for shift handover, root-cause analysis, and the compliance record — see shift handover and audit and compliance.