Transport intervals calibrated on the states the coupled runs visit
The transport intervals of entry 76 were calibrated on states sampled across the whole input box. On the states that TORAX runs actually pass through, they were wider than needed. The owner chose to add a second calibration on those states for the uncertainty work of M5 (OQ-44). The first intervals stay as the M4 record.
The data are the held-out test trajectories of the v3 dataset: whole TORAX runs, eight for ITER and thirty-two for MAST-U, never used in training. We fixed the split rule in a commit before computing anything. The runs are grouped by scenario family (ITER) or equilibrium (MAST-U) and shuffled within each group with a fixed seed. They are then dealt alternately into a calibration half and a report half. The reverse direction, calibrating on the report half, is computed as a check.
| half-width vs the box-wide one | coverage of the other half | reverse direction | |
|---|---|---|---|
| MAST-U (16 runs per half) | 0.21–0.29 | 0.870–0.894 | 0.905–0.924 |
| ITER (4 runs per half) | 0.25–0.32 | 0.936–0.966 | 0.711–0.818 |
The intervals become three to five times narrower in the model’s signed-log units. In flux terms, they shrink from a factor of 9–60 on 1 + |Q| to a factor of 1.8–3.3 (calibrated on half of the held-out v3 trajectory runs, the 90% half-widths are 0.21–0.32 of the V7 values; the intervals span a factor 1.8–3.3 on 1 + abs(flux) instead of 9–60C-259).
For MAST-U, both directions land near nine in ten, from 0.870 to 0.924, with coverage of single runs between 0.75 and 0.97 (calibrated on 16 held-out runs and tested on the other 16, coverage is 0.870–0.894; the reverse direction gives 0.905–0.924; single runs 0.75–0.97C-260). These intervals can serve as M5’s surrogate-error terms, with the run-to-run spread stated beside them.
For ITER the two directions disagree (calibrated on 4 held-out runs, the intervals cover 0.936–0.966 of the other 4; calibrated the other way they cover 0.711–0.818; not reliable for a new ITER runC-261). Calibrated on one half, the intervals cover 0.936–0.966 of the other half; calibrated the other way, only 0.711–0.818. The surrogate’s error differs between runs more than four runs per half can average out, so the result depends on which runs did the calibrating. Conformal prediction promises coverage only when calibration and test points are exchangeable, and here whole runs differ. The ITER loop-state interval is therefore not reliable for a new run. A reliable one needs more held-out trajectories, or a method that treats runs as the unit.
What’s next
M5 will need either more ITER test trajectories or a run-level calibration before it relies on an ITER loop-state interval. The remaining M4 steps are output bounds, local robustness, the EC closed loop and guard integration, followed by the certification review.
Technical details → Verification › conformal loop
Decisions → OQ-44