Every EC local robustness box crosses the training boundary
The next M4 steps are executed. The transport output proof remains unresolved, and the EC robustness assessment found a domain limitation that comes before the neural bound calculation.
| step | retained result |
|---|---|
| transport intermediate ranges | both selected transport members remain physically unknown; largest baseline activation widths fall to 0.246 and 0.449 of the preceding pilot, but inverse signed-log overflow persistsC-274 |
| exact EC state recording | 30 upper-launcher states (29 served) and 27 equatorial-launcher states (16 served) retain exact features, profiles, outputs, model identity, configurations and guard decisionsC-275 |
| EC local robustness | all 57 complete EC local boxes are out of domain, including all 45 served states; zero robustness proofs or model counterexamples are claimedC-276 |
| later reference comparison | twenty output states per launcher are selected at evenly spaced indices for the later GENRAY comparison; no reference solver execution or accuracy resultC-278 |
The transport experiment intersected intermediate backward bounds with interval propagation and used ELU’s monotone activation range. This reduced the largest intermediate ranges, but the physical inverse transform still overflows. The final optimized bounds are not uniformly tighter than the preceding pilot, and the small pilot was not expanded.
For EC, we ran the existing upper and equatorial launcher scenarios once each and recorded every accepted state’s exact surrogate inputs. The audit reconstructs them from saved profiles and checks model predictions and guard decisions.
The local robustness property uses symmetric perturbations of every continuous input, while fixing the launcher flag. The loop states sit near the lower effective-charge boundary, and both launch positions are at training-box edges. Their perturbations therefore leave the declared domain. This affects even states served by the inference guard: serving a point does not establish a neighbourhood around it.
The property forbids clipping. We retained every out-of-domain box and issued no robustness pass or model counterexample. OQ-45 asks the owner whether to carry this limitation or approve a distinct perturbation scope. The original result remains either way.
An initial package omitted dataset provenance sidecars. Its proof processes stopped during loading; those failures are saved. The corrected package passed the actual relocated loaders and reused the successful EC traces, without repeating the simulations. Focused validation is 70 focused checks pass normally and with Tier B disabled; lint, formatting and typing of all 92 source modules passC-277.
Next comes the concrete GENRAY input and compute package for the selected states. V5 needs a stronger bounded experiment; EC V6 needs its scope disposition. The MAST-U reference limitation remains unchanged, and the certification review is still pending.
Technical details → Verification › ec local robustness · Verification › intermediate range pilot
Decisions → OQ-22 · OQ-24 · OQ-26 · OQ-33 · OQ-42 · OQ-45