Validation boundaries

This page records current numerical and evidence limitations. It is a public description of observed behavior, not an operational backlog or work order.

Equilibrium source consistency

The bounded equilibrium lane computes the source-balanced Grad-Shafranov residual

\[ \frac{\lVert\Delta^*\psi + \mu_0 R J_\phi\rVert} {\max(\lVert\mu_0 R J_\phi\rVert,\Delta\psi)}. \]

The earlier source-free proxy misclassified shaped and high-beta equilibria by omitting the reconstructed \(R p' + FF'/(\mu_0 R)\) source. The current validator reconstructs \(J_\phi\) from GEQDSK \(p'\) and \(FF'\) profiles.

The available repository inputs have distinct evidence classes:

• SPARC GEQDSK inputs are public_reference design equilibria, not measured facility shots.
• DIII-D-like GEQDSK fixtures are synthetic and exercise numerical plumbing.
• A q95 value read from the same GEQDSK is a self-consistency observation, not an independent reconstruction comparison.

These inputs can pass a computational threshold without admitting physics, measured-shot, facility, public, or production claims.

Disruption-prediction replay

The repository replay data are synthetic. Recall and false-positive-rate results therefore describe the fixed-weight heuristic on those fixtures only. They are not a facility-database ROC, prospective warning-time study, or commissioned disruption-mitigation result.

Malformed or unsafe NPZ payloads fail closed per file and are never loaded with pickle enabled.

Transport scaling

The transport lane compares the implemented IPB98(y,2) calculation with curated published-reference rows. Its evidence class is public_reference. Agreement within the declared uncertainty band is a computational comparison against those rows; it does not establish a new multi-machine experimental validation.

Campaign-level interpretation

validation/validate_real_shots.py now emits the schema scpn-control.reference-evidence-validation.v1. Every lane declares exactly one of real, public_reference, synthetic, or local_proxy and reports data provenance separately from computational success.

The repository campaign mixes public-reference and synthetic evidence. Its real-shot, facility, public-claim, and production admissions are therefore fail-closed even when every numerical lane passes. The JSON and Markdown reports display those fields independently so a green calculation cannot be read as a green facility claim.

Reproduction

python validation/validate_real_shots.py --help
python validation/validate_real_shots.py \
  --output-json artifacts/reference_evidence_validation.json \
  --output-markdown artifacts/reference_evidence_validation.md

The validation guide explains the wider evidence taxonomy and the generated physics traceability report identifies component-level claim status.

Multi-machine validation

The compatibility facade remains validation.multi_machine_validation. Defining owners are validation.machine_inputs (inputs and presets), validation.synthetic_diagnostics (legacy noisy examples), validation.confinement_reference (local scaling comparison) and validation.equilibrium_execution (actual fixed-boundary execution). Existing defined-name imports and numerical statements are preserved; defining module metadata now identifies these owners. Native examples/renderers and the ordinary documentation gate inspect the defining owners directly.

validation.multi_machine_validation.MultiMachineValidator currently refuses run_all, save_json and save_markdown with RuntimeError. Its former seven passing metrics were generated from random values without evaluating the machine configuration or physical models. Those legacy outputs cannot support claims of equilibrium convergence, transport agreement, conservation, stability or diagnostic reconstruction.

The current refusal occurs before generating results or opening output files. Real machine-bound model execution and reference evidence are required before this API can publish validation results. Machine presets and the separately available synthetic diagnostic examples do not satisfy that requirement.

Capturing machine inputs

MachineConfig.snapshot(rho) captures the actual scalar parameters and sampled ne, Te and Ti profiles in versioned JSON with explicit units. Hash its UTF-8 bytes to identify the sampled inputs for a future model run. This identifies values; it does not identify callback source code or attest measured provenance. The radius grid must increase strictly from zero to one with at least three points. Invalid geometry, nonfinite values, negative profiles and mismatched profile shapes are rejected. Zero edge values are retained; a solver requiring positive edge values needs an explicitly supplied boundary contract.

Callbacks receive separate copies of the grid, and each output is captured before the next callback executes. The returned string remains unchanged if callbacks or machine objects are subsequently modified. Stateful callbacks can produce different snapshots on later calls; execute the eventual model against the captured inputs rather than resampling them. No equilibrium, transport, conservation, stability or reconstruction result is produced by input capture. The full seven-domain campaign remains unavailable pending its model and reference integrations. Executable examples live in the native docstrings of MachineConfig.snapshot and ConfinementReference.evaluate; the owning tests run them against the actual public APIs and shipped reference files. Computing a digest of these local files binds the example inputs but supplies no independent provenance or trust anchor.

Confinement reference comparison

ConfinementReference.evaluate() executes the existing IPB98(y,2) implementation on one explicit CSV operating point. The caller supplies the reference and coefficient paths with expected SHA-256 digests, zero-based data-row index, machine/shot identity, source classification and positive relative-error tolerance. Python and NumPy booleans are rejected as tolerances. Each file is read once, and those exact bytes are hashed and parsed. Changed bytes, ambiguous CSV headers, mismatched identities and invalid numeric inputs are rejected. Coefficient JSON also rejects duplicate keys at any depth and nonfinite numeric tokens, including overflow and unused metadata. Hash agreement alone does not establish a unique, valid model interpretation.

The returned JSON includes the source row, digests, input units, prediction and reference in seconds, relative error and the declared tolerance. It takes loss power and isotope mass explicitly from the row; preset auxiliary power is not a substitute. comparison_pass reports only the declared numerical comparison. Source classification remains caller-declared, training-domain membership is not assessed, and action/facility authority remains false. Supported source classes are derived_calibration, design_reference and synthetic; none establishes held-out empirical validation.

With the shipped coefficient/reference bytes and a 10% tolerance, the ITER design row predicts about 3.606 s against 3.70 s and passes the comparison. JET row 92436 predicts about 0.520 s against 0.41 s and fails. These are regression observations on the supplied illustrative reference data. They do not validate arbitrary profiles under the same machine name or complete the seven-domain campaign.

Shipped equilibrium configuration names

A filename or reactor label containing Validated does not establish a successful validation. An unchanged run of validation/iter_validated_config.json on the current Python SOR solver reached its 1,000-iteration limit with converged=false and an interior GS residual RMS of approximately 2.6191. Independent evaluation of the discrete operator on the returned flux/current arrays reproduced that residual. The configuration uses permeability 1.0 and current in configured units; these values cannot silently be interpreted as SI or matched to a 15 MA preset by name. Its rectangular computational domain also does not establish a measured plasma boundary or minor radius.

The solver normalises its current distribution toward the declared total. Integrating that same distribution back to the target checks that normalisation; it does not independently establish physical current conservation. A campaign must retain the convergence failure, explicit unit convention, boundary/source contract and separate comparison reference.

Capturing an equilibrium execution

EquilibriumExecution.evaluate(scratch_directory) runs the existing fixed-boundary solver on a temporary byte-identical copy of a SHA-256-bound configuration. The reactor identity must match, and a positive GS RMS acceptance threshold is supplied explicitly. Source files remain untouched; the temporary copy is removed after execution.

A configuration digest binds bytes; it does not certify every numerical solver control. The adapter refuses nonfinite returned fields/grids or independently computed GS RMS with RuntimeError, before returning JSON. Actual runtime regressions exercise malformed and extreme copied relaxation controls; a loose stopping policy in those negative copies grants no admission and changes no canonical solver policy. These tests retain the 1000-iteration budget and verify source preservation and temporary-copy cleanup. Solver/parser errors propagate.

The returned JSON retains original and effective configuration, solver histories, flux/current fields and R/Z grids. It independently applies the discrete GS operator to the returned fields and compares the residual against the declared bound. Numerical acceptance also requires the solver's own convergence result. Nonconvergence remains a recorded failure. Units remain configuration-defined; no SI conversion, plasma geometry or physical identity is inferred from names. This execution does not yet compare an external reference equilibrium, and its external-reference, action and facility flags remain false. Full seven-domain integration remains unfinished.

Status of the transport formulation

The subsections from here to the end of this page describe the transport discretisation implemented in this repository: density-weighted Crank–Nicolson diffusion, balanced internal exchange, conservative thermal storage and the prescribed face stage. It is a candidate numerical formulation. It has not been independently reviewed, and nothing on this page is a claim of physical validity, of validation against reference data, or that it is the reference transport mathematics.

Transport energy gate and returned profiles

The transport energy check now runs after internal exchange, pedestal overrides and final profile sanitisation. Previously, enforce_conservation=True could accept a state whose temperatures were subsequently changed before return, while energy_balance_error still described the earlier state. The corrected gate checks the final thermal profiles with the existing 1% threshold. The subsequent balanced-exchange correction removes the legacy per-channel rescaling. This does not establish an independent physical energy-balance reference or validate all transport approximations.

The initial energy also retains the pre-species density together with the initial temperatures. Using the evolved density with old temperatures previously rewrote the starting energy during a multi-ion step. The corrected accounting preserves a consistent initial state. Historical before/after measurements predate the balanced-exchange correction below; multi-ion rejection remains tested. This remains an internal check under the current ni = ne energy convention, not validation of a complete species-resolved physical energy model.

Public thermal accounting snapshot

TransportSolver.last_energy_balance exposes an immutable ThermalEnergyBalance after thermal assessment, including a step rejected by enforce_conservation. It contains timestep in seconds, auxiliary power in MW, initial/final/source energies in joules and the relative balance error. Consumers can check the reported arithmetic without reading private solver attributes. A saved record retains its values if profiles or the solver are subsequently changed.

Each evolve_profiles call first invalidates the previous record. A zero-time call or an error before assessment leaves None, so an earlier record cannot be mistaken for a measurement of that attempt. Nonfinite arithmetic may remain in a diagnostic record and must not be admitted as finite evidence. The record uses species-count ion capacity and electron density for its two channels; it is not a complete species-resolved or independent plant energy balance.

Balanced internal ion–electron exchange

Previously, electron equilibration had no opposite ion transfer and was counted as an external source. A retained 50 kJ auxiliary-input case reported about 101.8 MJ of net sources. This was a model defect, not a physical validation.

The step now transfers equal and opposite energies after the separate transport solves, using frozen-rate relaxation weighted by the respective ion and electron heat capacities. Internal exchange contributes zero to net external sources. The fixed edge temperatures remain boundary conditions; pedestal overrides and sanitisation still precede the final energy check. Per-channel zero-heating rescaling was removed because a recipient channel can legitimately heat through internal transfer with no auxiliary input.

Runtime regressions exercise hot ions, hot electrons, equal initial temperatures and small/stiff timesteps, checking core heat, bounded relaxation, source signs and fixed edges. The inherited equilibration rate is not independently validated. The common-ion-temperature closure, splitting accuracy and external physical references remain open; numerical boundary power is accounted separately. These changes do not complete seven-domain multi-machine validation or establish facility readiness.

Radiation power and thermal heat capacity

Bremsstrahlung and multi-ion tungsten radiation are supplied in W/m³. Their conversion to keV/s now divides by the channel heat capacity 3/2 n e_keV, consistent with the thermal energy used in the balance record. Previously the conversion omitted 3/2, so integrating the temperature sink removed 1.5 times the declared radiative energy. The same correction applies to the evolved profiles and the integrated source record. The existing half/half allocation of multi-ion line radiation between the channels is retained.

Public runtime tests independently integrate the prescribed radiative powers and compare them with the source record, both with and without auxiliary power, in single-ion and multi-ion modes. They supplement core-temperature tests, so correcting only the diagnostic cannot satisfy the checks. This fixes a units conversion; it does not validate the radiation fits or their coefficients, species-resolved heat capacities or the legacy single-ion impurity cooling proxy, whose output is already treated as a temperature rate.

The thermal capacity convention is consistent with the heat-equation prefactor in the TORAX equation summary. Full boundary, species and independent physical validation remain open.

Boundary constraints in the implicit transport solve

The axis Neumann condition and fixed edge temperatures are now imposed in the Crank–Nicolson linear systems before solving. Previously identity boundary rows solved for source-updated boundary values, and those values were overwritten afterwards. Neighbouring interior cells therefore used boundary values different from the returned profiles.

Public-runtime regressions reconstruct face fluxes from initial and returned profiles and check the interior Crank–Nicolson balance, including cells next to both boundaries. Equal channel diffusivities make internal exchange cancel in the sum equation. Four timestep/diffusivity combinations reproduce the original failure and pass with the boundary rows enforced. This verifies a discrete boundary contract. The numerical boundary-power record is described below; physical boundary calibration and general splitting accuracy remain unfinished. The existing final-state 1% energy gate has not been relaxed or corrected with an inferred boundary loss.

Density-weighted thermal diffusion

Transport now uses the conservative operator L(T) = (1/a²)/(rho n) d/drho(rho n chi dT/drho), with arithmetic face averages of n chi and cell density in the heat-capacity denominator. The explicit and implicit halves use the same density frozen after species evolution. Previously both operators diffused temperature with unit capacity while the energy record weighted temperatures by density; internal face transfers could then create or remove total heat when density varied.

Public-runtime tests independently compare the interior energy change with the two boundary face powers and radiative losses for rising and falling density profiles. Equal channel diffusivities let the internal exchange cancel in that comparison. A separate closed-boundary public CN solve checks total heat conservation with variable diffusivity and density. These checks do not replace independent validation of species evolution or unequal-channel splitting.

The generic radial operators accept an optional positive density vector. Omitting it retains unit-capacity diffusion for callers solving that equation; the thermal transport caller always supplies its density. Nonfinite, nonpositive, complex, Boolean and wrongly shaped density arrays are rejected. A uniform change in density units leaves the temperature operator unchanged. Here chi denotes thermal diffusivity; no new physical calibration of transport closures is claimed. The boundary-power record below uses this operator; multispecies heat capacities and physical boundary validation remain open.

Boundary terms in the public thermal energy record

ThermalEnergyBalance now reports two signed boundary contributions in joules:

• diffusive_boundary_energy_j: net heat entering the evolved interior cells through their inner and outer faces. It integrates density-weighted face conductivities and the old/new Crank–Nicolson temperature midpoint, separately for each channel, before internal exchange or pedestal overrides.
• prescribed_edge_energy_j: the energy needed to force the stored edge node to its fixed temperature, minus the volumetric source that the fixed boundary row does not evolve. This term follows the current nodal-volume convention; it is not an independently measured separatrix power.

Positive values add heat. source_energy_j still contains the original all-cell volumetric source integral. The relative error subtracts that source and both boundary terms from the final-minus-initial energy, with the same denominator and 1% admission threshold. Boundary terms are calculated from the actual operator and its imposed edge condition, not fitted to the observed residual. Previously, physically transported heat could incorrectly trip the source-only gate. Unmodelled pedestal energy, species changes and numerical recovery remain visible in the residual; adding boundary terms does not admit them automatically.

Runtime regressions cover inward/outward heat transport, rising/falling density, initial edge mismatch and both enforcement modes. Independent dense CN solves check the flux record with unequal, spatially varying channel diffusivities. A real pedestal override still fails admission when its energy is unaccounted. These checks establish discrete accounting under ni = ne, not complete plasma, multispecies, nonlinear splitting or facility validation.

Physical geometry in species diffusion

D/T/He diffusion now uses the cylindrical operator (1/a^2)/rho * d/drho(rho * D * dn/drho), with diffusivity in m²/s and physical minor radius in metres. The explicit stability step scales with (a * drho)^2; zero diffusivity uses a single source step. The species API requires rho and a_minor, and the integrated solver passes its actual grid and radius. Geometry-free callers must migrate to these explicit arguments. Invalid grids, radii, diffusivities and timesteps raise ValueError.

Public runtime tests compare parabolic-density evolution with the analytic cylindrical operator across three physical radii. Small-radius cases check the density maximum principle under diffusion and nonnegative burn losses. These verify numerical geometry, not a complete multispecies plasma model. Particle diagnostics account for signed boundary exchange as described below; numerical clipping remains an error and thermal source coupling remains incomplete.

Species-count thermal capacity

The multi-ion thermal solver now uses ni = n_D + n_T + n_He + n_impurity for ion heat capacity, assuming all ions share Ti. Electron heat capacity still uses ne; charge multiplicity does not create additional thermal ions. The public ion_density property returns a separate array. Single-ion mode retains ni = ne.

Stored energy, confinement time, pressure mapping, ion heating/radiation conversion, density-weighted diffusion and boundary power use these channel capacities consistently. Frozen-rate exchange preserves ni*Ti + ne*Te with unequal capacities; its inherited relaxation-time coefficient remains uncalibrated for a species-resolved collision closure. External source energy still excludes internal exchange, and the 1 percent admission threshold is unchanged. Changes in species populations are not automatically counted as external heat; their transport and reaction energy remain an open coupling contract. The explicit volumetric pumping heat model is described below. This is a thermal-capacity correction, not full multispecies energy conservation or validated fusion self-heating.

Integrated fusion consumption and helium pumping

Each species substep now applies diffusion, an analytic D-T burn update at fixed incoming temperature, then exponential helium pumping. Burn preserves n_D - n_T, consumes equal D/T counts and produces one He nucleus per reaction. It cannot consume more than the available limiting fuel. Pumping removes only available helium; positive infinite tau_He disables it.

SpeciesEvolutionResult.fusion_reactions and helium_pumped contain the integrated local counts in units of 10^19 m^-3 before imposed boundaries. S_He now means integrated production divided by the timestep, rather than the incoming instantaneous production rate. At zero timestep the three input species profiles remain unchanged and all integrated source diagnostics are zero. Callers relying on the previous instantaneous meaning must migrate. The particle residual uses actual integrated reaction and pumping counts, plus the independently computed boundary contributions described below.

Positive fuel floors were removed from the species kernel and thermal runtime; a prescribed D/T recycling density of 0.01 remains at the edge. Thermal evolution rejects a cell with zero ion heat capacity. Invalid negative/nonfinite species inputs and invalid pumping times are rejected by the species API.

Separate source operators are exact at fixed temperature, but their combined splitting is first-order. Repeated public calls demonstrate convergence against an independently integrated coupled burn/pumping reference. This does not calibrate the inherited fusion-reactivity fit or couple reaction and particle transport energy to the thermal solver. Pumping heat is specified below. Fusion self-heating and the full seven-domain machine validation remain incomplete.

Species boundary inventory

SpeciesEvolutionResult.particle_balance is a frozen scalar SpeciesParticleBalance record. Its initial/final counts measure D, T and He nuclei, rather than charge or density. One fusion reaction decreases this total by one; pumping removes the integrated helium count. Impurity inventory is not part of this D/T/He diagnostic.

diffusive_boundary_count integrates signed flux through the two faces bounding the evolved interior. prescribed_boundary_count measures the signed change from imposing the edge recycling densities after each substep. Positive values enter the domain. Both are computed from the applied operators, independently of the final inventory residual. The error is the absolute value of final - initial + fusion + pumped - diffusive_boundary - prescribed_boundary, divided by the initial count (denominator floor 1e-10). numerical_correction_count reports particles introduced by diffusion clipping; it is deliberately not subtracted from that error.

The supported quadrature requires dV = C * rho * drho, with finite positive C, zero axis weight, and positive interior and edge weights. Inconsistent weights raise ValueError; this diagnostic does not support arbitrary geometry.

TransportSolver.last_particle_balance exposes the completed species stage. Every thermal evolution attempt resets it before validating the timestep. Single-ion evolution and zero timesteps leave no record. A subsequent thermal admission failure preserves the completed species record for inspection; later thermal sanitisation or external profile mutation does not rewrite it. It is not a certificate for the final thermal-state inventory.

Public tests check signed inward/outward parabolic flux against an analytic reference, imposed-edge counts, combined diffusion/burn/pumping closure across physical radii and substeps, rejected volume weights, and record lifetime after real thermal rejection. Particle closure does not establish thermal coupling, calibrated fusion rates, or full machine validation.

Quasineutral electron capacity at low density

The multi-ion species result now retains ne = n_D + n_T + 2n_He + 10n_impurity without adding a minimum electron population. Runtime electron sanitisation no longer imposes the single-ion density floor or ceiling in multi-ion mode. The prescribed fuel edge therefore has electron density 0.02 in units of 10^19 m^-3 when no impurity is present, rather than the former artificial 0.1. A vacuum species result may have zero electrons; thermal evolution requires positive ion and electron capacities.

Auxiliary deposition rates are converted from the helper's floored reference capacity to the actual electron capacity before ion-capacity conversion. Electron radiation rates likewise use the actual positive capacity. Dilute public runtime tests reconstruct the initial stored energy and external source power from independent volume and radiation integrals. The inherited exchange time regularisation remains unchanged and uncalibrated. These changes remove an artificial charge/energy contribution; they do not supply the still-missing thermal energy transport or reaction closure; pumping heat is specified below.

Pressure consistency in the opt-in bootstrap approximation

The legacy bootstrap approximation now differentiates the same species-count pressure used by thermal storage and equilibrium profile mapping: P = (niTi + neTe) * 10^19 * e_keV, with the exact keV-to-joule conversion. The previous ne*(Ti+Te) expression overcounted helium ion pressure. Public bootstrap calls with analytic parabolic profiles check zero, moderate and high helium fractions against their pressure gradients.

This correction applies to the explicitly enabled legacy approximation. Its trapped-particle coefficient, current-sign convention and fixed effective-charge correction remain inherited and uncalibrated. It does not validate the default Sauter closure, the current-conservation domain, or particle-energy coupling.

Conservative thermal storage with evolving density

The thermal Crank–Nicolson step now discretises the change of nT. For each channel its storage is n_newT_new - n_old*T_old; the explicit conductive flux uses n_old and T_old, and the implicit conductive flux uses n_new and T_new. Volumetric heat sources retain the current channel-capacity normalisation. The thermal boundary record averages the corresponding old and new face fluxes, and prescribed-edge energy uses both old and new edge capacities.

Previously, changing density at frozen temperature created or removed thermal energy even when particle boundary flux, reactions and pumping were all zero. Six public runtime cases now close that energy balance, with rising, falling and uniform temperature and two timesteps. Independent dense matrix solutions use separate old/new storage and conductivity matrices to check local channel energy and boundary records. Single-ion fixed-density behaviour is retained.

This corrects storage for the existing heat equation with conductive transport and its declared sources. No separate particle-associated heat convection, pressure-work closure or fusion-product heating has been inferred from density changes. Without a specified heat source, changing population alone does not remove modelled channel energy: temperatures respond to the changed capacity. The ash-removal source is specified separately below. Discrete energy closure does not validate a physical source model. The remaining source/flux closures and coupled timestep convergence remain required for full multispecies validation. Numerical recovery and unmodeled pedestal changes remain visible to the unchanged 1 percent admission threshold.

Thermal energy carried by helium pumping

The distributed ash sink now uses a velocity-independent removal model: removed He nuclei carry their local mean ion thermal energy, with two accompanying electrons carrying their mean electron thermal energy. For the integrated local removed density p (in 10^19 m^-3), the step removes 1.510^19e_keVp(Ti_old + 2*Te_old) joules per cubic metre. The ion and electron parts enter their own channel equations using the completed positive capacities. Counts come from the actual substepped species kernel, including ash produced before pumping. Temperatures are frozen at entry to the outer step; combined thermal/source accuracy remains limited by that splitting.

ThermalEnergyBalance.helium_pumping_energy_j records this signed contribution. It is already included in source_energy_j and must not be summed again. The prescribed thermal edge row cancels ignored volumetric sources as before. Zero timesteps clear the step record; disabled pumping and single-ion mode contribute zero pumping heat. Independent no-fusion exponential-removal tests reconstruct the count-weighted ion/electron energy integral. A separate dense thermal oracle uses public species counts to check local deposition and channel profiles.

This is a volumetric thermal-ash model, not a flowing-plasma enthalpy or sheath model. It assumes He is thermalised at the common ion temperature and electron removal maintains quasineutrality; energy-selective pumping, fast-alpha dynamics, pressure work and boundary exhaust temperatures are not represented. Fusion product heating, particle-associated heat convection and full coupled convergence remain open.

Coupled local pumping, exchange and radiation convergence

A public runtime refinement test compares the uniform core against independently integrated two-temperature ODEs with exact exponential helium survival. The reference includes the inherited exchange-time coefficient, the runtime's arithmetic effective charge (including its prescribed edge node), and bremsstrahlung. With local mean-energy pumping, the removal terms cancel the corresponding density derivatives in the temperature equations.

At 64, 128 and 256 steps over 0.03 seconds, the maximum temperature errors are approximately 1.342e-4, 6.384e-5 and 3.160e-5 keV for initially hotter ions, and 8.657e-4, 4.304e-4 and 2.146e-4 keV for initially hotter electrons. Both sequences show first-order refinement while particle survival and per-step discrete energy remain near numerical precision.

This checks temporal convergence of the specified local thermal-ash model. Particle diffusion is disabled and the sampled uniform core is outside the thermal boundary layer. It does not establish spatial convergence, reactive D-T evolution, particle heat convection, or calibration of the inherited exchange and radiation coefficients. Those broader coupled checks remain open.

Preserve resolved gyrokinetic transport channels

The transport update now retains separate electron thermal diffusivity and D_n outputs from the quasilinear gyrokinetic, native TGLF and external GK paths. The ion channel retains its configured Chang-Hinton contribution. Previously, the final shared assignment overwrote both thermal channels with the ion-based estimate and replaced the computed particle coefficient by 0.1*chi_e.

Resolved turbulent channels no longer receive the additional legacy critical-gradient estimate. Gyro-Bohm/explicit constant-fallback modes retain that shared-channel prescription; existing H-mode pedestal profile handling remains separate. Real built-in quasilinear model comparisons at two electron/ ion temperature ratios verify retained channel outputs through the public transport update. This is integration correctness, not gyrokinetic calibration.

Native/external nested fallback is cellwise and uses the same three-channel composition: chi_gB = max(raw_gyro_bohm, 0.01), then chi_i = chi_gB + chi_nc, chi_e = chi_gB, D_n = 0.1*chi_gB on each failing non-core cell. It does not restore the direct legacy mode's additional critical-gradient term. The core/vacuum driver floor is 0.01 for each returned channel, before adding the ion neoclassical contribution. Both explicit fallback permissions are required. External execution errors are eligible for this fallback; native exceptions propagate, while invalid/unconverged native output can fall back. A selected transport mode is not evidence of actual backend success.

Public update tests compare a real native solve at rho=0.5 against independently specified local gradients at two electron/ion temperature ratios. Additional public tests use actual invalid native output and an actual absent external binary to verify refusal and permitted fallback composition. Configured real solver instances are supplied through the existing cached-solver attributes; no provider outputs are mocked. Malformed-config and missing-binary cases are routing robustness evidence. External successful execution and flux conversion remain unqualified; these tests do not establish physical calibration.

A real GACODE standalone run succeeds independently of this adapter. Its standard output is out.tglf.gbflux: signed particle, energy, momentum and exchange fluxes in gyro-Bohm normalisation, ordered with electrons first. These are not the chi_i, chi_e and d_e coefficients expected by the adapter's out.tglf.transport parser. A reference run exhibits inward particle transport; clipping that signed flux or labelling it a positive diffusivity would change its physical meaning. A flux-aware coupling and normalisation contract is required before the external path can be admitted.

The standalone signed-flux reader preserves normalised signed fluxes and multi-mode spectra from retained GACODE output and checks a present execution receipt. The launcher that executes actual decks in fresh directories with hashed execution evidence is a validation command of the repository, outside the installed package, because it needs an installed provider. Neither infers diffusivities or supplies the outstanding conservative runtime coupling.

The current adapter also emits a Fortran namelist where the standalone parser requires key=value records, supplies unsupported SHAT/ALPHA_MHD keys, and uses initialisation-only -i with a file path instead of executing a simulation directory. Real eigenvalue output has separate ky and multi-mode spectra, not the three-column table currently assumed. Correcting executable discovery alone therefore cannot qualify this external path.

D_n is still distinct from the configured D_species used by the D/T/He kernel. Mapping an electron-particle coefficient to species diffusion and deriving a particle-associated heat flux require an explicit closure; preserving the computed coefficient does not establish either mapping. No such closure is silently inferred by this change.

Radially varying species diffusivity

D_species accepts a finite nonnegative scalar or a real vector matching rho, within the finite float64 working range. Wider floating-point inputs above that range are rejected before conversion, face fluxes or CFL arithmetic. The common D/T/He operator is (1/a²)/rho * d/drho(rhoD(rho)dn/drho), using arithmetic averages of D at cell faces. Its signed boundary inventory uses the same face coefficients. The explicit physical CFL bound is 0.4(adrho)²/max(D); an all-zero profile retains the single source substep. Scalar input remains supported and constant-profile trajectories are identical.

Manufactured quadratic density/diffusivity cases independently check the cylindrical increment for rising/falling D at two radii, including the finite spacing term from arithmetic face averaging. A sharp alternating coefficient checks the maximum principle and zero numerical clipping across many substeps. Malformed, negative, nonfinite, boolean and complex profiles are rejected. A public thermal/species call verifies profile routing and input preservation.

All three fuel/ash species still share this configured coefficient. D_n is not automatically mapped into it: species-dependent transport, impurity motion, ambipolar closure and particle-associated heat convection remain separate physical requirements. This change supports spatial coefficients without claiming those model choices or calibrated machine transport.

Conservative prescribed face transport

The signed face-flux stage advances explicit electron, D/T/He particle number and electron/total-ion thermal energy through the public TransportSolver.evolve_fluxes method. Local updates and boundary ledgers use identical physical face areas and fluxes. Nonambipolar or nonphysical steps reject without profile mutation. This supplies conservative coupling for prescribed physical fluxes, including values from the explicit TGLF SI conversion. Automatic spatial provider sampling, normalisation and species mapping remain unqualified; source/operator composition is explicit.