The MAST-U candidate replay fails reference coverage
The owner approved the prepared local replay once. The solver completed with finite, converged profiles, but the comparison stopped because a face served by the candidate used fallback in the saved physics reference: the single approved replay finishes in 65 seconds within its 240-second bound, with 40 finite converged steps and 146 passing integrity checks; comparison exits with a reference-coverage failure at candidate-served face 4, normalized radius 0.16C-189.
That reference cannot establish the formal temperature criterion for this candidate. The failing face stays in its actual served mask. We neither intersect masks nor substitute the historical baseline mask for acceptance. The saved trajectory, source snapshot, provenance and failure logs remain available; no solver retry was made.
The supplementary comparisons retain mixed results: candidate/baseline serve 15/13 faces with 11 overlapping; fixed-baseline-mask candidate T_e/T_i errors are 42.1252%/48.1182%, versus baseline 35.3331%/49.1458%; candidate whole-profile errors are 29.7365%/41.8856%, and own-mask descriptive distances to the hybrid reference are 29.8865%/40.0063%C-190. Electron temperature worsens on the fixed baseline mask. Smaller whole-profile errors and lower descriptive distances on the candidate’s own faces do not establish improved coupled accuracy. Those own-mask distances compare against a trajectory containing fallback at the uncovered face and are explicitly ineligible for acceptance.
Raw guard reconstruction matches every saved accepted-state face. Rejections decrease: candidate rejects 415/1040 accepted-state face records (39.9038%), versus baseline 466/1040 (44.8077%); exclusive candidate reasons are 341 model-box exits, 50 excessive diffusivities and 24 excessive convection values, with exact raw guard reconstructionC-191. This is accepted-state accounting; it does not measure intermediate solver calls, and the owner has not set a fallback-rate acceptance threshold.
The historical reference lacks raw rejection reasons. Repeating identical reference compute would not establish that the missing face becomes admissible. Its cause and a physically justified remedy need review before a new reference request. A separate sampling experiment can also be prepared from training-only evidence, with original holdouts and the physical contract preserved. Neither experiment has been authorized.
K-030 remains open. The current model and guard stay selected, and nothing has been synchronized or replaced.
What’s next. Review the reference-coverage failure and bounded follow-up proposal before new training or coupled compute. Formal certification remains unfinished.
Technical details → Verification › candidate replay result
Decisions → OQ-30