The v3 surrogates meet the held-out target
Dataset v3 samples where TORAX trajectories go, with whole runs kept out of training (entry 22). The final ensembles have now been trained on the broad v1/v2 data plus v3’s training points. The monotone variants meet the owner’s held-out accuracy target on both machines: MAST-U: Q_e 17.8%, Q_i 25.2%; ITER: Q_e 7.2%, Q_i 10.8%; all meet the 30% targetC-129.
An audit loaded the saved models, checked their metadata and dataset hashes, replayed the trajectory split, and recomputed the scores. The v3 test set was never a training input. This checks the implementation’s reproducibility; it is not independent physics validation.
Structure and its cost
Electron and ion heat fluxes are nonnegative in use. Each member is built so that its heat flux cannot fall as its own temperature gradient rises, with the other inputs fixed. Recomputed checks found no negative member slopes on the tested v1/v3 points or scan grids, and no negative neighbour differences on the scan gridsC-130. Those sampled checks support the construction; proof over the whole input domain belongs to M4.
TGLF itself is not always monotone. On the broad scans where its flux decreases, the final models depart from it: MAST-U: Q_e 3.14, Q_i 3.36; ITER: Q_e 2.51, Q_i 2.11 signed-log unitsC-131. This includes approximation error as well as the cost of imposing monotonicity. Passing the v3 median target does not remove that ceiling or establish accuracy everywhere in the training box.
Clipping also does not identify every stable point. The exact-zero rates and uncalibrated spread coverage are mean ±2 standard deviations covers Q_e/Q_i at 75.3%/74.4% (MAST-U), 85.6%/80.8% (ITER); exact-zero predictions where TGLF is nonpositive: 44.0%/47.0% (MAST-U), 52.3%/21.1% (ITER)C-137. Calibration and the guard’s uncertainty policy remain M4 work.
Inside the loop
The comparison command now selects the final monotone v3 model explicitly. A local RAPTOR L-mode run finished with all profiles finite; no stalled or coarse-tolerance step; post-step guard rejection 33.1%; execution 1.62 s after compilationC-135. Against the saved TGLF run at the same final time, its profile differences are T_e 7.72%, T_i 9.23%, n_e 11.36%, psi 0.32%C-134. Electron and ion temperature agreement improved relative to v1; density agreement did not. This is one case, not a closed-loop bound.
The strict derivative and discrete-balance checks exposed a stopping-precision limitation: v3 derivative discrepancy about 1.8e-5 relative; peak discrete energy residual 1.11e-5 (v1) and 1.36e-6 (v3)C-140. Converging Newton more tightly, to residual_tol=1e-10, gives v1 and v3 pass: three-step derivative FD tolerance 1e-7, forward/reverse tolerance 1e-12; guard masks unchanged over the stencilC-138 and v1 and v3 pass: energy residual below 1e-6 and particle residual below 1e-10, with fine solver convergenceC-139. The original assertion limits are retained. The runner records the solver setting, and the default-tolerance limitation remains K-026.
A second execution check, on ITER hybrid, gives 50 steps to 5 s, all four profiles finite, no stalled or coarse-tolerance steps; accepted-state guard rejection 26.9%C-142. It checks coupled execution and guard behavior, without a TGLF error bound.
The completed earlier timing comparison used v1. A fresh matched v3 comparison was submitted, but the scheduler cancelled it before execution and produced no output. There is no matched v3 speed-up measurement to report. The job script is prepared for a rerun once that compute issue is resolved.
M3 remains in progress: the held-out transport target passes, while matched v3 timing, broader coupled comparisons, and the RF and pedestal surrogate tasks remain open.
Technical details → Surrogates: final v3 results. Decisions → OQ-19, OQ-20.
Technical details → Surrogates › v3 results · Surrogates › in the loop
Decisions → OQ-19 · OQ-20