Decisions

Questions the agent is not allowed to settle, and how the owner settled them

The brief lists the decisions that belong to a human: physics modelling choices, licence and data questions, which properties to prove, use of the compute cluster, and benchmark failures that aren’t explained by a bug. When the agent reaches one, it writes the question down with context and a recommended answer, carries on with whatever isn’t blocked, and waits. Numbers (OQ-n) match the project’s internal record. Newest first within each group.

Awaiting a decision

OQ-32 · Reference treatment for signed heat transport

The diagnostic identifies negative raw heat diffusivity with valid solver inputs (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). Its physical/numerical origin and a physically justified reference treatment need review. Keep the current guard and reference admissibility criterion while reviewing the source/input evidence. New collection, remedies and matched acceptance compute need separate concrete reviews.

OQ-17 · Package licence metadata

What licence text the package metadata should carry while no release licence is chosen remains open. It blocks nothing while the code is private.

Settled

OQ-31 · Reference diagnostic compute

Decided 11 Oct 2026. Approve the exact prepared diagnostic once. It completes and reproduces the historical profiles/mask (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); signed heat transport explains the uncovered face (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). The diagnostic approval is consumed. It approves no reference remedy, fitting, new candidate replay, synchronization or replacement. Formal R6 acceptance remains not evaluable.

OQ-30 · Candidate coupled replay

Decided 11 Oct 2026. Approve the exact bounded local CPU replay once after its source and reference preparation. Its solver completes, but the comparison refuses saved-reference coverage at a candidate-served face (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). The attempt is consumed; no retry or new reference compute is approved. K-030 and the current model/guard remain in place, with follow-up experiments and replacement requiring review.

OQ-29 · MAST-U candidate training compute

Decided 11 Oct 2026. Approve the prepared bounded candidate-training experiment after split/lineage checks and the training-only smoke (entry 46). Synchronization, new coupled reference compute and model replacement remain separate reviews. The single approved job completed and passes its integrity audit (entry 47); local offline checks are recorded, with coupled revalidation pending.

OQ-28 · Private refresh artifact synchronization

Decided 11 Oct 2026. Approve the reviewed private upload for the completed refresh. Remote artifact checksums and downloaded bytes match (all 26 uploaded artifact records covering 18 distinct files match both remote SHA-256 metadata and downloaded bytes; the folder and private visibility match the approved request, and the offline payload queue is emptyC-184), and the run remains private. This approval covers synchronization; training remains separate work.

OQ-27 · Bounded MAST-U remediation compute

Decided 11 Oct 2026. Approve one bounded training-only feedback label refresh after the failure diagnosis, keeping both test equilibria excluded. The approval covers label collection; training and model or guard replacement remain separate work.

OQ-26 · Floating-point export parity

Decided 11 Oct 2026. Keep the relative allowance and add an explicit absolute allowance in each declared output unit (including signed-log training space), after diagnosing cancellation near zero. The selected criterion passes: 13 export checks pass in each mode: all selected members, random box points, local holdouts, physical transforms and served means, with rtol=1e-12 and atol=1e-12 in each declared output unitC-180. No physics or closed-loop tolerance changed.

OQ-25 · MAST-U remediation first

Decided 11 Oct 2026. Prioritize diagnosing and remediating the failed MAST-U transport comparison (served-face T_e 35.33%, T_i 49.15%, both failing 15%; whole profile 41.51/62.39%; cells with both adjacent faces served give 11.86/23.24%, still failing T_i; 13 of 26 final-state faces servedC-179), with the chosen limit unchanged. A new compute request requires separate approval after a concrete remedy is prepared.

OQ-24 · Verification engine

Decided 11 Oct 2026. Use auto_LiRPA with input-box splitting, keeping a patched alpha,beta-CROWN package as a fallback if local-robustness proofs are too slow. Certificates will name the actual engine and pinned version. Gurobi is not installed.

OQ-23 · Model storage for CI

Decided 11 Oct 2026. Store the selected model versions, explicit input boxes and their certificates with the code. CI will re-prove the quick properties and fail on mismatched model/certificate digests. The training datasets are not required to read the stored boxes.

OQ-22 · Properties, coverage and tolerances for verification

Decided 11 Oct 2026. M4 proves properties of the surrogates over input boxes, calibrates their uncertainty and compares closed-loop trajectories with the physics codes; the brief leaves the choice of properties to the owner.

Decision. The owner accepted the agent’s recommendations. Transport: heat flux non-decreasing in its temperature gradient, certified output bounds over the training box, and local robustness on sub-boxes around the loop’s states. EC: certified bounds of current, centres and widths per launcher, and the heating centre moving by at most 0.05 for input changes of one percent of their ranges. ITER pedestal: certified bounds of pressure and width around the ITER states TORAX visits. Calibrated intervals with 90% coverage. Closed-loop tolerances: transport within 15% of TGLF on the radii the surrogate serves, EC within 10% of GENRAY on sampled loop states. The original verification tool was alpha,beta-CROWN on checked PyTorch copies; OQ-24 selects auto_LiRPA with input splitting instead.

OQ-21 · RF and pedestal references

Decided 10 Oct 2026. Reference, model and channel choices were recorded on 8 Oct: GENRAY for EC heating and current drive, and the pretrained EPEDNN network for the ITER pedestal. The prescribed MAST-U pedestal remains, with predictive closure deferred; IC is deferred. A bounded numerical pilot of the original ITER deck completed (the corrected EC reference completes in 83.16 s with readable native profilesC-156).

Question (8 Oct). How do the reference outputs become labels and TORAX inputs: the current-drive method, heating channels, the current definition and its sign, the radial mapping, the pedestal branch, temperature split, density, pedestal-top location and the fallbacks; and what to label first.

Decision (10 Oct). The owner accepted the agent’s recommendations as written. EC: the deck’s adjoint current drive, with the momentum-conserving method as a check on a few cases; electron heating from both channels, kept separately; current as GENRAY’s flux-surface average of \(j \cdot B\) over \(B_0\), the same \(B_0\) as TORAX’s, signed positive along the plasma current; a remap that keeps every cell’s heating and current integrals; labels per MW injected; the Gaussian EC model fitted to the labels as fallback; targets after a baseline. Pedestal: both diamagnetic corrections, solution 0; equal ion and electron temperatures; the prescribed density; top at one minus twice the half-width in poloidal flux, mapped through the equilibrium; the prescribed pedestal as fallback. The first EC dataset samples ITER states from the transport dataset’s TORAX runs, with two launcher families (entry 36).

Targets and K-028 (10 Oct, later). The owner accepted the recommended EC targets: absorbed power and driven current within a median relative error of one tenth, heating centre within 0.05 in radius, on held-out runs. For K-028 the owner chose to drop states whose current is far from the equilibrium’s, and approved compute for the structured model. The structured EC surrogate meets all three targets (entry 38).

OQ-20 · Monotonicity where TGLF is not monotone; the ITER collisionality range

Decided 3 Oct 2026. OQ-19 asked for monotonicity in the gradients to be measured on TGLF before it is imposed. On TGLF scans of \(a/L_{T_e}\) and \(a/L_{T_i}\) around v1 test points, the heat flux falls along most scans (falls somewhere in 77% (Q_e) and 80% (Q_i) of MAST-U scans, 51% and 74% of ITER scans; by more than half in 40% and 57% (MAST-U), 34% and 57% (ITER)C-113), and the v1 ensemble’s slope is negative at a minority of test points (entry 20). Separately, the approved ITER collisionality range excluded part of the ITER loop’s radii from dataset v3, and the guard rejects them (K-023). Two questions: (1) drop monotonicity as a requirement, or impose it although TGLF is not monotone; (2) widen the ITER collisionality range.

Decision. (1) Monotonicity is a requirement for the TGLF surrogates: \(Q_e\) non-decreasing in \(a/L_{T_e}\), \(Q_i\) non-decreasing in \(a/L_{T_i}\). Where TGLF decreases, the surrogate departs from TGLF by construction; the departure is measured and reported as a ceiling, not hidden. (2) The ITER collisionality range is widened for dataset v3 only; the M2 sampling space and the v1 and v2 datasets are unchanged. The owner also approved the GPU retraining of the structured variants and the cluster job that compares the surrogate with TGLF inside TORAX.

OQ-19 · What to sample next, the accuracy target, and the sign of the heat flux

Decided 3 Oct 2026. On the M2 datasets, a uniform box in 17 inputs, the surrogates’ error falls slowly with more data, and the first round of active learning gained nothing over uniform points (entry 19). The configuration had no accuracy target. The brief asks for positivity, zero flux below threshold and monotonicity in the gradients, while TGLF’s own heat fluxes are negative in a minority of cases (Q_e < 0 in 2.0-6.5% of cases (below -1 gB in 0.5-1.5%; median of the negative values -0.03 to -0.09 gB); Q_i < 0 in 9.4-14.1% (below -1 gB in 2.2-5.1%); particle fluxes negative in 33-54%C-102). Three questions: (1) keep sampling the box, or sample elsewhere; (2) which held-out target; (3) a ReLU on the heat-flux outputs (the owner’s question).

Decision. The owner left (1) and (2) to the agent’s recommendation, which stands, and asked for (3). (1) Dataset v3 per machine is sampled where the coupled simulation goes: TGLF inputs from TORAX states along trajectories over a family of scenarios per machine, with the gradients and kinetic inputs perturbed around each state and the geometry from the same equilibrium as the profiles (this lifts K-018 for v3). MAST-U profiles are simulated with TORAX on FreeGSNKE MAST-U equilibria; no experimental data are used. v1 and v2 stay as the broad set. The TGLF labelling job is approved as usual. (2) Median relative error of the ensemble mean at most 0.30 on Q_e and on Q_i, on cases with |Q| above 1 gyro-Bohm unit, on each machine’s v3 held-out split. Errors on the v1 split are reported, not gated; the closed-loop error is checked at M4. (3) Q_e and Q_i get a ReLU at the output: exactly zero below threshold, never negative. The particle fluxes keep their sign. Training must not let the ReLU block gradients for points predicted stable (a linear output in training with a penalty on positive predictions where TGLF’s flux is not positive, clipped in use, as QLKNN does). Setting TGLF’s negative heat fluxes to zero is a cost, to be listed as a ceiling. Monotonicity is measured on TGLF scans before it is imposed.

OQ-18 · Close M2 without active learning, QuaLiKiz and the gated-code drivers?

Decided 3 Oct 2026. M2’s three acceptance criteria were met on the two version 1 datasets, but three of its tasks were not built: active learning near the critical gradients, a QuaLiKiz driver, and drivers for the gated (Tier B) codes. Decision. (a) Active learning moves to M3: it needs a model of where the thresholds are, and the M3 ensembles provide one; the new points become dataset version 2, v1 unchanged. (b) QuaLiKiz is deferred: its shifted-circle geometry does not represent the MAST-U shapes, and QLKNN already serves the conventional case. (c) A driver interface with stubs for the gated codes, which refuse to run without access, was built. M2 closed (M2 page). For M3 the owner set: train on every case, with the fluxes in a sign-preserving log space and no filtering; one ensemble per machine; training on GPUs, each cluster job approved; the accuracy target to be set after a first baseline (entry 18).

OQ-16 · Kinetic ranges, ITER shapes and the α_MHD cap for the M2 datasets

Decided 2 Oct 2026. The owner set one dataset per machine, MAST-U and ITER. The geometry ranges were measured (OQ-15); the kinetic ranges, the ion species and a cap on the normalised pressure gradient α_MHD were proposed by the agent from memory and needed a decision before any data was generated. Three questions: (a) do the kinetic ranges and ions stand per machine; (b) is one ITER plasma shape with the current ramp-up’s q profiles enough, or should 15 MA baseline equilibria be added first; (c) a fixed cap, a cap that depends on the magnetic shear, or sampling α_MHD in place of β_e.

Decision. (a) Approved as proposed for both machines: D and carbon for MAST-U, D-T as one ion of mean mass and neon for ITER (ranges in datasets). (b) One shape is enough for version 1; the missing baseline is recorded as a limit (K-016) and baseline shapes come before an ITER version 2. (c) Fixed caps for version 1 (MAST-U 4, ITER 2), a shear-dependent cap to follow (K-017). The two datasets were generated the same day (entry 17).

OQ-15 · Sampling ranges and TGLF settings for MAST-U

Decided 2 Oct 2026. MAST-U is outside both surrogates TORAX ships with (aspect ratio fixed at 3, or elongation up to 1.5). The agent proposed ranges from memory, not checked against a source. Decision: the agent measures the geometry ranges on equilibria rather than choosing them, and the kinetic ranges go to a physics check (became OQ-16); a second region besides MAST-U, each tagged so it can be selected alone or pooled (became the ITER dataset); TGLF with the SAT2 saturation rule and both magnetic perturbations; CGYRO spot checks of TGLF at M4. Measured, the MAST-U geometry is narrower in shape than the proposal (triangularity on interior surfaces at most 0.24, not 0.6).

OQ-14 · One cluster job to check TGLF and time it

Decided 2 Oct 2026. TGLF’s licence was settled by the owner’s pointer to the public GACODE repository (Apache-2.0). The remaining install check, TGLF as TORAX’s transport model, could not run on a login node. Approved: one CPU node for at most 30 minutes, running TORAX’s TGLF tests, a TORAX step with TGLF, CGYRO’s regression suite, and TGLF on a few hundred points of the sampling space to measure the cost per case. Outcome (entry 16): TGLF 9/9 regression cases; tglf2py wrapper 11/11; TORAX TGLF transport-model tests 3/3; CGYRO 21/21C-069; 2.2-2.5 s per case alone, 2.6 s median in a 64-worker pool; about 700 core-hours per 10^6 casesC-068; the TORAX step with TGLF failed (fails at the first step: NaN fluxes on the axis face, about 10^6 gyro-Bohm units on the edge face (K-015)C-070, K-015). The first submission failed at start-up on a shell setting in the job script, with no compute used, and was resubmitted.

M1 physicist review

Passed 1 Oct 2026. The last M1 acceptance criterion is a review by a physicist. The owner did the review as reviewer of record. The agent set out six items, each with its evidence and a recommendation; the owner accepted all six recommendations as written.

1. Benchmark tolerance, against RAPTOR on the published ITER L-mode case (OQ-12): the t = 10 s profiles within 2% for T_i, T_e, ψ and q, and within 0.1% for the densities; the line-averaged temperature and density traces within 5% over 0–10 s. The fusion and Ohmic powers are reported, not gated: the paper’s two codes differ by 13–22% on the alpha ion/electron split. Result at the published settings: T_i 1.23% (1.1%), T_e 0.28% (0.71%), n_e 0.03% (0.03%), psi 0.49% (0.41%), q 0.69% (0.70%)C-050; transient 2.4% at dt = 0.05 s, 2.9% at dt = 0.00625 s (1.0% at t = 10 s)C-053. Both pass.
2. TORAX version. The differences between the paper’s TORAX run and TORAX 1.4.3 (T_i 0.13%, T_e 0.50%, n_e 0.01%, psi 0.08%, q 0.16%; Ohmic power density 3.6%C-051) are accepted as a code-version effect.
3. Pedestal. The fixed-pedestal model’s pinning source removes 87–98%C-038 of the heating in the ITER hybrid ramp-up. Accepted for core profiles inside ρ = 0.9; power crossing the edge is not taken from these runs (K-010). Revisited at M3.
4. Time step. The ITER hybrid example settings are not converged in time (entry 3). They are used only to check the wrapper. Runs used as a physics reference use a step at which halving it changes the core temperature by at most 1% after the first 5 s. Solver stalls stay reported (OQ-5).
5. ASCOT5 alpha source. Orbit tolerance 10^-7 (0.348 at 10^-6, 0.331 at 10^-7C-058 of the birth power to ions at 10^-6 and 10^-7); 640 markers give the total (72.2 MW vs 70.5 MWC-059 against TORAX’s local model) and the electron/ion split, and the radial shape waits for the production run (OQ-13). Holding the source fixed for 2 s is accepted (+0.03% in stored energy (central T_e +0.5%, T_e and T_i profiles 0.6%)C-060). The small excess in the markers’ birth power (K-011) is to be explained before neutral-beam ions are added.
6. Fallback, pedestal and sawtooth models: OQ-1 below.

The claims these items judged are marked reviewed in the claims register.

OQ-1 · The fallback transport model

Decided 1 Oct 2026, at the M1 physicist review. M0 needed a toy transport model to exercise the loop, and a smooth critical-gradient form was used. The trusted fallback for real runs is TORAX’s Bohm/gyro-Bohm model, which is already the fallback behind the guard inside TORAX. The critical-gradient form stays as the closure of the Tokamak Toolkit’s own analytic test loop only. For M1 the pedestal is TORAX’s fixed pedestal, and sawtooth mixing is represented by a constant diffusivity near the axis (the ITER examples have no sawtooth model); a sawtooth model is added when a case needs one.

OQ-13 · A third ASCOT5 job, for the production run

Decided 1 Oct 2026. The second ASCOT5 job (entry 12) finished its tolerance check but not its production run: a per-marker time limit, set too short by the agent, stopped 2823 of 6400 markers, 18% of the birth powerC-057. A third job is approved: up to eight hours on one node, the production run at the tighter tolerance with no time limit, and a check at a tighter one still. The M1 physicist review is not held for it.

Outcome, 2 Oct 2026 (entry 15). The job ran in 3 h 37 min and every alpha finished. Total heating: 71.8 MW vs 70.5 MWC-064. The tighter check changed the heating by electrons -0.16%, ions +0.08%C-066, so the working tolerance stands. No further job is requested.

OQ-12 · Which published profiles the benchmark is compared with

Decided 1 Oct 2026. M1 asks for a match to published profiles. The ITER hybrid scenarios have no published profiles from a second code. The TORAX paper publishes one quantitative comparison with an independent code, RAPTOR, on an ITER L-mode case, with its full configuration and the profiles as vector figures. That case is the M1 physics benchmark, compared with both RAPTOR’s and the paper’s own TORAX curves, first at the published time step and then at a converged one (entry 11). The ITER hybrid runs remain the stiff-transport and pedestal case. The tolerance was set at the M1 physicist review.

OQ-11 · A second ASCOT5 cluster job

Decided 1 Oct 2026. The first ASCOT5 job (entry 10) showed that the orbit tolerance chosen for speed loses alphas that should be confined (C-043), and that at the tighter tolerance 640 markers resolve the total alpha heating but not its radial shape (C-045). A second job is approved: one node for up to four hours, ten times more markers at the tighter tolerance, and a check at a tighter one still. Further jobs need their own approval. Outcome: entry 12 and OQ-13.

OQ-10 · Which radial variable the guard’s features use

Deferred 1 Oct 2026, to M3. The guard checks whether a plasma state is inside the region the surrogate was trained on, using features such as normalised temperature gradients. The Tokamak Toolkit computes those gradients with respect to the normalised flux radius, scaled by the geometric minor radius; QLKNN, as TORAX feeds it, uses the midplane minor radius of each surface. On the ITER hybrid scenario the two differ from +19%C-032 in the core to about a third at the edge (K-008). No run is affected yet. The convention will be chosen when the first trained surrogate exists (M3), and in any case before a verified domain is certified (M4).

OQ-9 · Fast-ion sources from ASCOT5

Decided 1 Oct 2026. Fusion alphas first, born from the plasma profiles TORAX computes, as a direct check against TORAX’s own alpha heating. Neutral-beam ions follow once a physicist has approved the injector geometry. One short test job on the compute cluster is approved; each further job needs its own approval.

OQ-5 · What “reproduces the benchmark” means

Decided 1 Oct 2026. Two comparisons: the Tokamak Toolkit’s loop against TORAX’s own output at the same settings (a check of the wrapper), and against the published cross-code comparison (a physics statement). The wrapper check must agree to 10^-10 (it passes at round-off: C-006, C-007). The physics tolerance is set by the physicist reviewer at the M1 gate (set: see the M1 physicist review). Because the published reference runs are not converged in time (entry 3, C-012, C-013), the physics comparison is made at the published settings first, then at a converged time step, with the difference reported as its own line in the discrepancy budget. Solver stalls stay reported, not tuned away by relaxing a line-search setting (entry 5, C-011). Energy and particles must balance to 10^-6 per step on converged steps (entry 8, C-035, C-036); the flat-top reference solver’s defect (C-037) is reported, not required to close.

OQ-3 · Where the code will live

Decided 1 Oct 2026. A private hosted repository, where the continuous-integration workflow runs on every change. The code stays private; this log is how the work is shown. Until the workflow has passed, “tested” in the claims register still means “asserted by a named test, run before each publish”.

OQ-8 · Which machine description for the equilibrium solver

Decided 30 Sep 2026. MAST-U. FreeGSNKE’s repository already ships a MAST-U-like machine description (coils, passive structures, limiter, wall; credited to UKAEA) under the same open licence, so nothing had to be assembled from published drawings. It is fetched from a pinned upstream version with a checksum per file. This covers the equilibrium work only: the M1 transport benchmark stays with the ITER scenario, because no published TORAX MAST-U scenario exists and QLKNN was not trained on spherical-tokamak conditions.

OQ-7 · Input data for a STEP extension

Deferred 30 Sep 2026. Extending the benchmark to a STEP power-plant flat-top is a future extension, not part of M1. The work stays with the ITER benchmark and MAST-U for now; the data question reopens if and when the extension is picked up.

OQ-6 · Which benchmark

Decided 21 Sep 2026, amended 30 Sep 2026. The ITER hybrid scenario that ships with TORAX: an 80-second current ramp-up, and a 5-second flat-top as the quicker check. A planned STEP extension was deferred on 30 Sep (OQ-7).

OQ-4 · Newton-Raphson for the coupling loop

Decided 21 Sep 2026. Approved. The M0 loop used relaxed Picard iteration, which oscillates instead of converging once transport becomes stiff: on the toy model’s original settings the residual sat between 3 × 10⁻² and 8 × 10⁻² after 12 iterations. The owner approved a Newton-Raphson solver with derivatives from automatic differentiation and a fixed iteration count, keeping Picard as the fallback, and stall reporting for both (entry 1).

OQ-2 · Where to run

Decided 21 Sep 2026. Tests and the early milestones run on an interactive node with one datacentre GPU. Anything longer goes through the cluster’s batch system on an allocation the owner named, and every job is announced to the owner before it is submitted.

The corrected original ITER numerical pilot completes with readable native heating and signed-current profiles (the corrected EC reference completes in 83.16 s with readable native profilesC-156). Native accounting and descriptive saved-reference comparisons are retained; the rendered PostScript page is blank. Production EC coupling contracts, grouped sampling and ITER boundary choices remain open.