PlasmaGuard learns disruption risk from real FAIR-MAST tokamak shots, quantizes it in a Rust replay engine, and turns it into auditable CONTINUE / MITIGATE / ABORT decisions — uncertainty-aware and fail-closed.
A tokamak disruption can damage a device that costs a fortune and months to rebuild. An ML model can estimate disruption risk, but a floating-point probability is not a safety decision — and a model that's confidently wrong under uncertainty is worse than none. What's missing is a control layer that is deterministic, replayable, and biased toward safety.
A model estimates disruption risk with a conformal upper bound — uncertainty made explicit.
Risk is quantized from float to fixed-point bps, so the same shot always replays into the same chain.
If the conformal bound crosses the threshold, it moves toward MITIGATE or ABORT — never relaxes.
Built on public FAIR-MAST MAST tokamak data — not synthetic telemetry.
Under uncertainty the system tightens toward MITIGATE, never toward risk.
Every shot reproduces the exact same auditable decision chain.
Let's talk about deterministic, replayable, fail-closed decisioning.