MAST-U retraining preparation preserves the existing holdouts
The completed feedback-label collection now has a separate retraining configuration. Its source stays at the recorded location. An optional path mapping extends the loader, while existing configurations keep their previous paths and split rules.
The preparation freezes source hashes, original split indices and feedback-row lineage. It compares retained training rows with held-out rows using exact model-feature and common solver-input matches. The audit found no cross-split duplicates. The new source is training-only. Its case coordinates and equilibrium identities remain attached to the original full solver inputs.
The model’s feature definitions, signed-log output transform, heat-flux clipping, stability penalty and required monotonicity remain the same. Normalization will be recomputed from the augmented training rows. That expected difference is recorded; the declared physical input box stays fixed.
The focused checks and bounded local smoke pass: 55 focused checks pass in both normal and Tier A modes; the bounded training-only smoke produces finite predictions and reloads exactlyC-185. The smoke uses training rows for both fitting and its internal checkpoint fixture. Its output is finite and reloads exactly. These are tooling checks, not held-out accuracy results.
The dedicated batch script runs a single MAST-U candidate from a checksummed source snapshot. It records actual source sampling, selected checkpoint epochs, history and provenance. A concrete allocation request is ready for the owner. Production training has not run, and K-030 remains open.
What’s next. Review the compute request, then evaluate any approved candidate against the unchanged offline and coupled criteria before a model decision.
Technical details → Surrogates › m4 feedback preparation
Decisions → OQ-29