EC and pedestal output bounds; transport limits remain unresolved
The output-bound calculation now runs on the saved surrogate members. It checks each export against JAX, bounds the output across the declared input box and splits that box within an explicit budget. The results differ between the three surrogates.
| surrogate | result |
|---|---|
| EC | all five EC members have finite CROWN bounds for both launcher families; current and width bounds remain very looseC-262 |
| pedestal | all nine pedestal pressure and half-width pairs have finite bounds over the recorded ITER trajectory box, with the widening and domain clipping statedC-263 |
| transport | all twenty members of each transport ensemble have finite training-space bounds that overflow the inverse signed-log transform; no finite physical guard domain is certifiedC-264 |
For EC, each launcher family keeps every other input range from the stored training box. The serving clamps limit absorbed power, deposition centres and minimum widths. Bounds on current and maximum width remain much broader than the outputs sampled inside the box. Finite enclosure therefore does not yet mean a useful physical limit.
The old pedestal diagnostic saved its predictions but omitted the network inputs. The updated diagnostic records those inputs and the model hash. A fresh run supplied the visited span; the proof box widens it by a stated margin within the upstream domain. All pressure and half-width solutions are retained, including the solution used by TORAX. There is still no comparison with labelled EPED data.
For transport, the finite bounds in signed-log training space are so wide that the inverse transform overflows. The reports retain those bounds and mark the physical result unknown. This is a limitation of the bound calculation within its budget, not evidence that the network reaches those extremes or a reason to change the acceptance limits.
Each split covers the original domain completely. Sampled extrema check enclosure and measure tightness; they never substitute for proof. The audit binds the model, property, source and evidence, and checks the retained partitions. Production guards still use their existing contracts.
What’s next
Tighter transport bounds are needed before V5 can supply physical guard domains. The remaining M4 work also includes local robustness, the EC coupled comparison, guard integration and the owner’s final certification review. MAST-U’s accepted scope limitation remains in force.
Technical details → Verification › output bounds
Decisions → OQ-22 · OQ-24 · OQ-26 · OQ-42