EC loop-state power and current pass the fixed GENRAY reference criterion
M4 needs a reference check of the electron-cyclotron heating surrogate on states the coupled simulation actually visits. The saved state selection now has a complete GENRAY comparison, with every case retained.
The package reconstructs the model inputs from saved plasma profiles before rendering the solver decks. It retains the original state indices, times, launch settings and guard decisions. Source, model, traces, decks, reference equilibrium, executable and runtime libraries are bound by checksums. An exclusive attempt marker prevents a changed output path from rerunning the same request.
| item | evidence |
|---|---|
| frozen reference cases | forty decks bound to the unchanged selection, exact model and traces, frozen source, equilibrium, executable and runtime; eight workers with a thirty-minute hard limitC-279 |
| affected validation | 40 affected tests pass normally and with Tier B disabled; repository lint, formatting, typing of 93 modules and batch syntax passC-280 |
| completed reference comparison | 40/40 cases pass the unchanged 10% power and signed-current limits, including 31/31 served states; maximum current error 9.606% EL and 7.393% UL, maximum power error 0.07483% EL and 0.02599% ULC-281 |
The comparison includes both electron heating channels and preserves the current sign convention. It reports every case, separating states where the guard served the surrogate from rejected predictions. A rejected prediction is diagnostic: it was not the source used by the simulation. Missing or failed reference outputs prevent a complete comparison. A zero reference is handled explicitly, without introducing a denominator floor.
The reference keeps the existing fixed-equilibrium convention. Substituting new profiles into that equilibrium does not produce an equilibrium-matched reference trajectory. The package and result retain this limitation.
Tests check reconstruction, units, signed errors, failure accounting and evidence tampering. They use synthetic outputs and fake executables; physical agreement comes from the separate batch execution. The bounded request completed under the standing M4 compute authorization. The independent audit reconstructs native totals, reloads the model for fresh inference parity and checks every guard decision. Every selected case passes both criteria, including the served subset. The rejected predictions remain diagnostics. V9 is complete within its stated reference convention. Transport output bounds, the EC robustness scope and the certification review remain open.
Technical details → Verification › ec genray reference comparison
Decisions → OQ-21 · OQ-33 · OQ-42