A plan for completing M3 and starting verification
We have written a dependency-ordered action plan for the remaining surrogate work. It distinguishes completed local validation from performance measurements, reference choices and subsequent verification.
The local transport checks are complete: 15 gradient, JIT, fallback and discrete-balance checks pass across RAPTOR and ITER hybrid in each modeC-146. The final-model matched runtime comparison remains open because its compute run was cancelled before execution. The next performance step is to resolve that cancellation, agree a fresh bounded submission and compare the final surrogate with reference transport at matched settings. Newton profiling stays separate from the linear-solver timing comparison.
In parallel, the owner can choose the reference solvers or datasets for radio-frequency (RF) and pedestal responses. That decision must specify channels, outputs, units, criteria and Tier A fallbacks. The existing guards, grouped dataset interfaces and baseline trainer are ready for those references; no predictive RF or pedestal model is trained.
After reference selection, we label a small grouped baseline, inspect errors and physical integrals, agree accuracy targets, train ensembles and integrate them into TORAX behind guards. Gradient, balance, fallback and coupled cost checks then define completion.
M3 closes only when all its required surrogate tasks meet the accuracy, differentiable execution and speed-up criteria. Verification follows: agreed properties and trajectory tolerances, formal proofs, uncertainty calibration, reference comparisons and certification review. Local sampled checks remain distinct from those certificates.
This documentation update selects no new physics closure and submits no compute job. The build log now follows the same sequence as the internal plan.
Technical details → Next steps.
Technical details → Surrogates › next