The approved MAST-U candidate is trained and checked offline
The owner approved one bounded MAST-U training experiment. Its single submitted job completed within the approved allocation and retained each ensemble member’s selected checkpoint, along with the training and validation history.
The completion audit binds the saved model to the reviewed source snapshot and frozen data manifest. It verifies original split indices and feedback-row lineage, checks the actual sampling counts and compares the resulting normalization and physical box with the preparation. The saved parameters and tested predictions are finite: 164 completion checks pass, binding the saved candidate to the reviewed snapshot and frozen data/splits, with finite parameters/history and tested member/served predictionsC-186. The candidate stays in a separate directory.
The local offline comparison keeps validation, existing holdouts and refresh training states separate. Results are mixed: the existing held-out flux gates pass, one flux median improves and the other regresses slightly. Refresh training medians improve modestly while their tails worsen. Those training scores do not measure generalization, and passing the flux gates does not establish coupled temperature accuracy.
| check | recorded outcome |
|---|---|
| approved training | completed within its budget |
| source, split and checkpoint integrity | completion audit passes |
| numerical export parity | all 168 export comparison records pass the unchanged mixed tolerance, covering all 20 members in training/physical units, the served ensemble and JAX JIT/eager arithmeticC-187 |
| monotonicity | saved-parameter and sampled derivative diagnostics pass; formal proofs pending |
| coupled comparison | candidate replay pending; K-030 open |
Every exported member and the served ensemble pass the unchanged mixed numerical parity tolerance. Sampled physical derivatives have no negative values for the required monotone pairs. This checks the implementation and saved parameters; it does not replace the formal M4 proof.
The current model and guard remain selected. Tracking stays offline, and no new artifacts have been synchronized. The training approval did not include a coupled solver run.
What’s next. Prepare and review the matched candidate replay, retain its paired reference and served-mask diagnostics, and judge both temperatures against the existing limits before any model decision.
Technical details → Surrogates › m4 feedback candidate
Decisions → OQ-29