Action plan

M3 done; verification in progress

Current position

M3 closed on 10 October (entry 41). M4 now has owner-selected properties and tooling, validated property contracts, and a transport comparison passing on ITER and failing on MAST-U (entry 42). The exports pass the numeric criterion approved in OQ-26; MAST-U remediation takes priority (OQ-25). Formal proofs, calibration, EC reference comparisons and certificate-bound guards remain ahead. The certification review is pending. Matched fluxes and an exact accepted-state replay diagnose the MAST-U failure (entry 43). The approved training-only label refresh completed, and its local artifact checks pass (entry 44). Artifact synchronization is complete with downloaded checksums verified (entry 45). The approved candidate is trained and checked offline (entry 47). Numerical export checks pass, and paired metrics retain mixed regressions. Matched coupled revalidation remains ahead; its tooling and reference audit are prepared (entry 48). The approved attempt finishes its solver but fails saved-reference coverage (the single approved replay finishes in 65 seconds within its 240-second bound, with 40 finite converged steps and 146 passing integrity checks; comparison exits with a reference-coverage failure at candidate-served face 4, normalized radius 0.16C-189; entry 49); review reference admissibility and a bounded follow-up before new compute. Keep K-030 and the current model/guard.

The sections below retain the M3 work record.

M3 remains in progress. The final monotone transport ensembles meet the agreed held-out accuracy target (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). Local derivative, fallback and conservation checks cover RAPTOR L-mode and ITER hybrid (15 gradient, JIT, fallback and discrete-balance checks pass across RAPTOR and ITER hybrid in each modeC-146). The ITER hybrid derivative checks use an explicit Newton solver override; the existing linear-solver timing scenario is separate.

Guards, grouped response datasets and a baseline trainer are built for radio-frequency (RF) and pedestal responses. Their software checks use synthetic numerical fixtures (25 synthetic contract tests pass in full and Tier-A-only modesC-143, 26 synthetic baseline-pipeline tests pass in full and Tier-A-only modesC-145). No new RF or pedestal model is trained. The owner-selected pretrained ITER pedestal model now has a JAX port with numeric checks (12 upstream scalar regression values agree within relative tolerance 1e-12C-149, 12 focused tests pass in full and Tier-A-only modesC-150) and a conversion to TORAX’s pedestal fields (pedestal pressure 0.0772 MPa, T_ped 3.59 keV at n_e,ped 7e19 m^-3, top at psi_norm 0.935C-162); coupled acceptance remains unfinished. Final-model matched speed-up is measured (TGLF 539.1 s, final v3 surrogate 1.249 s, QLKNN 0.117 s; the v3 surrogate is about 432 times faster than TGLF and 10.7 times slower than QLKNNC-158). EC dataset v1 is labelled (3120 of 3120 cases produce readable output (20 of 20 check cases); median 84.8 s of GENRAY time per caseC-164) and has a first held-out baseline (driven current RMSE 3.30 kA/MW against 13.5 kA/MW for the training mean; pointwise profile model: current-density RMSE 2.20e4 A/m^2 against 2.09e4 A/m^2 for the meanC-166).

Complete the performance comparison

Resolve the cancelled compute run, then agree a fresh bounded submission. Compare TGLF, the final monotone surrogate and QLKNN using matched scenario, grid and solver settings. Record compilation separately from execution, resource allocation, finite profiles, solver status, profile errors and accepted-state guard rejection. Every output carries provenance.

Profile Newton separately to distinguish ensemble inference, transport/geometry work and Jacobian construction. Keep unresolved overlap visible (K-022). Historical speed-up belongs to the original model; timings from different hosts or drivers cannot establish final-model speed-up.

Done when: a reproducible final-model runtime comparison and discrepancy report exist, with allocation and solver settings stated.

Done 10 Oct. The cancellation matched a recurring node start-up failure (seven full-node jobs since 2 Oct were cancelled by root at start with no reason; each resubmission landed on another node and ranC-161); the approved resubmission ran and gives TGLF 539.1 s, final v3 surrogate 1.249 s, QLKNN 0.117 s; the v3 surrogate is about 432 times faster than TGLF and 10.7 times slower than QLKNNC-158, with profile deviations final v3 surrogate: T_e 7.7%, T_i 9.2%, n_e 11.4%, psi 0.32%; QLKNN: T_e 155%, T_i 128%, n_e 86%, psi 2.4%C-159 (entry 35). The runner now rejects coarse-tolerance results, records effective CPU allocation, separates warm-up from execution and records final-state guard diagnostics after timing. Focused checks pass (11 performance-tool checks pass in full and Tier-A-only modesC-147). The offline saved-trace audit records unresolved component attribution (72059 complete HLO events with about 5.369 s of covered trace timeC-148); K-022 remains open. Draft entry 28 is in the private review copy.

Complete the EC and ITER pedestal contracts

The owner has selected the reference, pedestal model and EC channels in OQ-21. EC labelling and the ITER pedestal port proceed independently. The prescribed MAST-U pedestal remains, with predictive closure deferred; IC is deferred.

Complete the EC launch/current-drive prescription, power convention and coordinate/units mapping. For ITER, choose the pedestal solution and map pressure/width to supplied density, temperature and boundary conditions. Retain explicit Tier A fallbacks and acceptance criteria.

The native EC output reader preserves separate channels and SI units; its focused checks pass (24 focused tests pass in full and Tier-A-only modesC-151). The original ITER deck is selected for the numerical pilot only; production launch and current-drive settings remain open.

The saved upstream ITER output also passes the reader/accounting check (200 native bins in the saved upstream ITER EC output pass reader and accounting checksC-152). The bounded pilot runner retains immutable inputs, outputs, logs and failures; its combined checks pass (40 focused pilot, reader and Tier A checks pass in full and Tier-A-only modesC-153). The external reference now builds with a replacement plotting library (one external reference executable builds with verified artifacts and resolved runtime librariesC-154). The corrected numerical pilot is complete with readable native profiles (the corrected EC reference completes in 83.16 s with readable native profilesC-156). Native accounting and descriptive saved-reference comparison are retained; the rendered PostScript page is blank. Settle the production coupling contracts and grouped sampling before reviewing a separate labelling budget.

Done when: those contracts and required access are recorded. Done 10 Oct: the owner accepted the recommended conventions (OQ-21, entry 36).

Train and integrate the response surrogates

Define physically related case groups and label a small baseline. Keep entire groups together in training and holdout sets. Inspect failed labels, sampled domains and physical integrals. After the baseline, agree held-out accuracy metrics and targets with the owner.

Train structured EC ensembles and evaluate untouched holdouts. The selected pretrained ITER pedestal is ported directly; no second network needs to learn its outputs. Integrate source and boundary responses into TORAX behind guards with Tier A fallbacks. Supply channel adapters required by the selected targets. Check gradients, finite execution, energy/particle balance, physical rejection, fallback use and coupled cost. Pointwise profile errors alone do not establish absorbed-power accuracy.

Done when: the reference datasets, checkpoints, accuracy results, guarded coupled checks and matched cost measurements have provenance and meet the agreed criteria.

EC dataset v1 is labelled (3120 of 3120 cases produce readable output (20 of 20 check cases); median 84.8 s of GENRAY time per caseC-164). The first baseline learns the totals but its pointwise profile model does not beat the training mean (driven current RMSE 3.30 kA/MW against 13.5 kA/MW for the training mean; pointwise profile model: current-density RMSE 2.20e4 A/m^2 against 2.09e4 A/m^2 for the meanC-166), so the EC model needs a structured profile output. K-028 records that every case uses one equilibrium. The owner then set the targets and restricted the states (K-028 option b); the structured surrogate meets all three targets (all three targets pass: absorbed power 0.01% and driven current 2.7% median relative error (targets 10%), heating centre 0.0036 median absolute error in rho (target 0.05)C-168, entry 38). Both the EC surrogate and the ITER pedestal now run inside TORAX behind their guards (runs under jit for both launchers with the surrogate serving the source; energy balance within 1e-6 and particles within 1e-10; derivatives through Newton match finite differences to 1e-7C-171, runs under jit with EPEDNN setting the pedestal; energy balance within 1e-6 and particles within 1e-10; derivatives through Newton match finite differences to 1e-7C-173, entry 39).

Close M3, then start M4

Audit the complete M3 criteria: held-out accuracy, differentiable guarded execution and measured end-to-end speed-up. Transport accuracy alone does not close the RF/pedestal tasks. Keep the ion-flux convention, monotonicity departure, default-root tolerance and domain limits visible (K-024, K-025, K-026).

After M3 acceptance, agree verification properties and trajectory tolerances with the owner. Prove properties over declared input boxes, calibrate uncertainty on held-out data, compare trajectories against Tier A references and prepare per-surrogate verified-domain reports. Certification review closes M4. Sampled monotonicity and raw ensemble spread are not certificates.

Keep the build log current

After each completed step, update its narrative entry, technical section, claims, ceilings, milestone state and executed statistics. Counts that reuse unchanged execution reports retain that qualification. Drafts and disclosure-held pages remain in the private review copy until the owner reviews them. The owner renders and publishes the separate website handoff.

M3 closed on 10 Oct 2026 (entry 41). The reference diagnostic is now consumed (the diagnostic completes in 517 seconds within its 30-minute bound, with 40 finite converged steps, exact historical temperature/density/poloidal-flux-profile and final-mask equality, and 171 passing saved-output integrity checksC-193). Immediate work: review the signed heat-transport evidence and reference remedy (OQ-32; final face 4 at normalized radius 0.16 has valid inputs but raw electron heat diffusivity -0.02792124888 m²/s; negative heat diffusivity rejects that face at 8 of 40 accepted states; reference rejection is 323/1040 accepted-state face records (31.0577%)C-194), preserving reference inadmissibility and the separately proposed sampling experiment. Continue M4 under its existing decisions (OQ-22). See M3 status, decisions and known ceilings.

Entries 21 to 34 and the active-learning and surrogates pages were approved on 9 Oct. New entries follow the same review before publication.