The MAST-U candidate replay preserves masks and guard accounting
The trained candidate now has a separate coupled-replay runner. An explicit immutable model path selects it for this experiment while the ordinary default keeps selecting the existing model. The source package binds the runner, configuration and inputs before execution.
The preparation audits the saved reference against its recorded runner and clean source revision. Scenario settings, core transport code and guard settings agree. The original accepted-state replay also reproduces the saved baseline exactly, letting us check the new reporting without another solver run. Historical external dependency bytes were not recorded; that limit remains explicit. A new run must still match actual grids, time points, settings and convergence before the reference can be reused.
The focused checks and a bounded training-only smoke pass: 40 focused checks pass in normal and Tier A modes; the two-step training-only candidate smoke is finite and reconstructs the raw accepted-state guard predicate exactlyC-188. The smoke uses a training equilibrium and two steps, with raw diagnostics after stepping. It is a tooling check, not the held-out coupled replay or an accuracy result.
| report | purpose |
|---|---|
| candidate served faces | unchanged formal temperature criterion |
| baseline mask and overlap | expose changes in the served region |
| fixed baseline mask | supplementary comparison on the same historical faces |
| whole profile and conservative cells | retain broader error diagnostics |
| raw accepted-state guard reasons | explain rejections without claiming solver-call accounting |
An isolated served face is retained. If no cell has both adjacent faces served, the cell metric is marked unavailable rather than discarding the face. Rejection reasons include both overlapping predicate counts and an exclusive priority partition. Candidate history must reconstruct the saved raw guard mask exactly; historical missing generic guard features are reported as a limitation.
A concrete local CPU replay request is ready for the owner. The held-out candidate replay has not run, and K-030 remains open. The current model and guard remain selected.
What’s next. Review the bounded replay request, then report both temperatures and all paired diagnostics against the existing limits before any model decision.
Technical details → Verification › candidate replay preparation