Verification Matrix

Purpose

This page is the research-facing index of what SPECTRAX-GK treats as verified, validated, exploratory, or deferred. It is meant to answer four questions for each lane:

1. what physical model is being exercised,

2. what observable is compared,

3. what the reference is,

4. what acceptance gate applies.

Literature Baselines Reviewed

The current matrix is anchored on these published baselines:

• Tronko et al., Verification of Gyrokinetic codes: theoretical background and applications: verification methodology, observed-order checks, and benchmark-observable framing.

• Mandell et al., GX: a GPU-native gyrokinetic turbulence code for tokamak and stellarator design: CBC, W7-X, KBM, nonlinear transport, velocity-space convergence, and performance figure conventions.

• González-Jerez et al., Electrostatic gyrokinetic simulations in W7-X geometry: W7-X ITG/TEM scans, zonal-flow response, and nonlinear ITG heat flux.

• Nevins et al., Characterizing electron temperature gradient turbulence: ETG operating-point conventions.

• Monreal et al., Residual zonal flows in tokamaks and stellarators at arbitrary wavelengths: residual-zonal-flow metrics and damping interpretation.

• Merlo et al., Linear multispecies gyrokinetic flux tube benchmarks in shaped tokamak plasmas: shaping scans, ballooning-angle handling, Rosenbluth-Hinton residuals, and GAM damping.

• González-Jerez et al., Electrostatic microturbulence in W7-X: comparison of local gyrokinetic simulations with Doppler reflectometry measurements: fluctuation amplitudes, frequency spectra, and zonal-flow spectral content.

• Maurer et al., Global electromagnetic turbulence simulations of W7-X-like plasmas with GENE-3D: heavy-electron electromagnetic verification before realistic-mass stellarator production runs.

Status Legend

• Closed: benchmark lane is accepted for research claims.

• Open: lane is active and expected to close.

• Exploratory: useful for development, not yet a paper claim.

• Deferred: intentionally out of scope for the current paper/release.

Tokamak Linear

Lane	Observable	Reference	Status	Baseline gate
Cyclone ITG	gamma(k_y), omega(k_y), eigenfunction overlap	GX + CBC literature	Closed	rtol <= 1e-2 except documented low-k_y / near-marginal cases
ETG	gamma(k_y), omega(k_y)	GX + ETG benchmark literature	Closed	rtol <= 1e-2 on the tracked branch
KBM	gamma(k_y), omega(k_y), branch continuity vs beta	GX	Closed	rtol <= 1e-2 on the accepted branch
KAW	branch-followed linear response	GX	Deferred	close branch identity before publication use
TEM	gamma(k_y), omega(k_y)	GX / literature	Open	close branch-following and reference selection first
Shaped multispecies tokamak	gamma, omega, eigenfunction shape	Sauter benchmark set	Open	literature-backed operating point and overlap gate required
Shaped tokamak zonal-flow / GAM	residual level, damping rate, GAM envelope	Merlo et al. + analytical Rosenbluth-Hinton estimates where applicable	Open	residual and damping must match literature/code-backed references before publication use; signed Phi_zonal_mode_kxt is now available. The current artifact is docs/_static/miller_zonal_response_pilot.png from tools/generate_miller_zonal_response_pilot.py using Merlo Case-III Table-III parameters, an initial density perturbation, a common pre-recurrence fit window t≈30, separate positive/negative-extrema damping fits, and a Hilbert-phase frequency extraction on that same window. It gives residual≈0.192 against a paper-scale target of about 0.19, ω_GAM R0 / v_i≈2.20 against a figure read-off near 2.24, and γ_GAM R0 / v_i≈-0.176 against a figure read-off near -0.17. The remaining explicit follow-up item is the later finite-moment recurrence rather than the benchmark-scale Merlo gate

Frozen artifact paths for the currently closed tokamak linear lanes:

• docs/_static/cyclone_comparison.png

• docs/_static/etg_comparison.png

• docs/_static/kbm_comparison.png

• docs/_static/kbm_eigenfunction_overlap_summary.png

• docs/_static/reference_modes/kbm_linear_gx_ky0p3000.npz

• docs/_static/benchmark_core_linear_atlas.png

Closed raw-overlay diagnostic artifacts for the KBM lane:

• docs/_static/reference_modes/kbm_linear_spectrax_ky0p3000.csv

• docs/_static/kbm_eigenfunction_reference_overlay_ky0p3000.png

• docs/_static/reference_modes/kbm_eigenfunction_reference_overlay_ky0p3000.json

• tools/generate_kbm_reference_overlay.py

The refreshed bounded-cost extraction produces normalized overlap 0.999985 and relative L^2 mismatch 0.00721 against the frozen GX raw mode at k_y \approx 0.3 when run with the exact KBM grid contract, the selected growth-fit window, and a late-time eigenfunction tail window. The generator writes a machine-readable gate report with overlap >= 0.95 and relative L^2 <= 0.25 as the acceptance policy, and this raw-overlay artifact now passes.

Branch-followed scan tables should use the same gate-report convention: observed-order gates for resolution or velocity-space convergence, and branch-continuity gates for adjacent gamma/omega jumps and successive eigenfunction overlap when overlap data are available. The tracked KBM candidate table now has a no-rerun summary path through tools/generate_kbm_branch_gate_summary.py and docs/_static/kbm_branch_gate_summary.json. That summary now uses the continuity-first selected branch and passes the strict checks: max_rel_gamma_jump ~= 0.388, max_rel_omega_jump ~= 0.320, and no successive-overlap deficit.

Observed-order convergence tables should also gate both the asymptotic finest refinement and the full set of pairwise refinement orders. The generic tools/generate_observed_order_gate.py path now records this policy in JSON. The tracked Cyclone velocity-space convergence artifact docs/_static/cyclone_resolution_observed_order.json is closed on an office/GPU ky=0.30 time-path sweep with all pairwise orders positive, final-pair order above 4.8, and finest-grid relative growth-rate error about 1.1e-3.

The current materialized gate reports are indexed by tools/make_validation_gate_index.py in docs/_static/validation_gate_index.json and docs/_static/validation_gate_index.png. Exploratory diagnostics can set gate_index_include=false so they remain documented but do not count as release blockers. The current release-gate index has 16/16 tracked reports passing.

Stellarator Linear

Lane	Observable	Reference	Status	Baseline gate
W7-X ITG flux tube	gamma(k_y), omega(k_y)	stella/GENE benchmark paper + GX	Closed	rtol <= 1e-2 on closed adiabatic-electron ITG branches. This row does not close W7-X TEM or kinetic-electron validation; those remain blocked by the explicit TEM branch audit below.
W7-X TEM / kinetic-electron extension	gamma(k_y), omega(k_y), multi-alpha/multi-surface windows	stella/GENE benchmark paper + W7-X TEM literature	Open	docs/_static/tem_branch_parity_audit.json is outside the publication parity envelope and docs/_static/w7x_tem_extension_status.json keeps multi-alpha, multi-surface, and kinetic-electron nonlinear windows open. Do not use the closed W7-X ITG row as a TEM claim.
W7-X zonal flow	residual level, damping envelope	stella/GENE benchmark paper + zonal-flow literature	Open; time coverage closed, residual and late-envelope gates open	a case-specific runtime/tool path exists through examples/benchmarks/runtime_w7x_zonal_response_vmec.toml and tools/generate_w7x_zonal_response_panel.py. The runtime now supports the paper-facing init_field="phi" Gaussian potential initializer and writes both the older volume-weighted Phi_zonal_mode_kxt diagnostic and the W7-X line-average Phi_zonal_line_kxt observable. The frozen VMEC-backed artifact now lives at docs/_static/w7x_zonal_response_panel.png with metadata in docs/_static/w7x_zonal_response_panel.json and replayable traces in docs/_static/w7x_zonal_response_panel.traces.csv; it uses line-first normalization, following the paper text. The stella/GENE Fig. 11 reference traces and inset residuals are digitized by tools/digitize_w7x_zonal_reference.py into docs/_static/w7x_zonal_reference_digitized.csv and docs/_static/w7x_zonal_reference_digitized_residuals.csv. tools/compare_w7x_zonal_reference.py writes the residual/time-coverage/envelope artifact docs/_static/w7x_zonal_reference_compare.json and can now replay the comparison from the tracked combined trace CSV. tools/plot_w7x_zonal_contract_audit.py writes docs/_static/w7x_zonal_contract_audit.png as a publication-facing open-lane diagnostic, tools/plot_w7x_zonal_moment_tail_audit.py writes docs/_static/w7x_zonal_moment_tail_audit.png to track the velocity-space recurrence / moment-tail hypothesis, tools/plot_w7x_zonal_closure_ladder.py writes docs/_static/w7x_zonal_closure_ladder_kx070.png to compare bounded closure attempts at k_x rho_i=0.07, tools/plot_w7x_zonal_state_convention_audit.py writes docs/_static/w7x_zonal_state_convention_audit.png to close the initializer/observable convention layer, and tools/plot_w7x_zonal_recurrence_sweep.py writes docs/_static/w7x_zonal_recurrence_sweep_kx070.png to separate moment-resolution and closure-source effects. The current SPECTRAX artifact enforces the intended test-4 k_x rho_i values [0.05, 0.07, 0.10, 0.30] with a periodic radial box for the k_y=0 zonal run and reaches t≈3460 for k_x rho_i=0.05 and t≈1980 for the other three wavelengths. Under the paper-facing normalization, residuals fail at k_x rho_i=0.07, 0.10, and 0.30 and late envelopes fail for all wavelengths. The state-level audit closes the convention question with Gaussian-profile relative L2 error 1.85e-6 and helper/manual observable agreement near 2e-16. The bounded recurrence sweep shows that Nl=12,Nm=48 has the best no-closure trace error on t v_t/a <= 100 and that constant-source closure suppresses the final Hermite tail but worsens the trace error. The remaining closure step is therefore a physical velocity-space recurrence / damping fix rather than a documentation or normalization change
W7-X fluctuation spectra	resolved k_y spectra, k_x-k_y fluctuation power, and temporal spectra	W7-X nonlinear gate plus Doppler-reflectometry comparison conventions	Initial simulation diagnostic closed; experimental transfer-function validation deferred	tools/plot_w7x_fluctuation_spectrum_panel.py writes docs/_static/w7x_fluctuation_spectrum_panel.png with CSV/JSON/PDF companions from the gated W7-X t≈200 nonlinear NetCDF output. The script refuses failed nonlinear gate summaries by default and marks the JSON with claim_level = "validated_nonlinear_simulation_spectrum_not_experimental_validation" and gate_index_include = false. It therefore closes the reproducible spectrum-estimator layer while leaving density/zonal-frequency comparison through a Doppler-reflectometry transfer function as a future manuscript extension
HSX	gamma(k_y), omega(k_y)	GX / internal frozen references	Closed	near-marginal deviations documented explicitly
Electromagnetic stellarator verification	heavy-electron linear/nonlinear EM response	GENE-3D verification conventions	Open	close heavy-electron EM lane before realistic-mass claims

Frozen artifact paths for the currently closed stellarator linear lanes:

• docs/_static/w7x_linear_t2_scan.csv

• docs/_static/hsx_linear_t2_scan.csv

• docs/_static/w7x_linear_t2_lastvalue.csv

• docs/_static/hsx_linear_t2_lastvalue.csv

• docs/_static/reference_modes/w7x_linear_gx_ky0p3000.npz

• docs/_static/reference_modes/w7x_linear_spectrax_ky0p3000.csv

• docs/_static/w7x_eigenfunction_reference_overlay_ky0p3000.png

• docs/_static/reference_modes/w7x_eigenfunction_reference_overlay_ky0p3000.json

• docs/_static/benchmark_core_linear_atlas.png

For W7-X, the whole-window scan and the late-time last-value reduction tell the same story. For HSX, the whole-window mean_rel_gamma metric is kept as an honest near-marginal stress signal, but the late-time closure should be read from docs/_static/hsx_linear_t2_lastvalue.csv because the final (gamma, omega) values are much tighter than the whole-window average.

The W7-X raw eigenfunction overlay is now closed at k_y rho_i = 0.3 using tools/generate_w7x_reference_overlay.py. The frozen GX bundle was refreshed from the finite t≈2 raw field history because the older bundle source contained non-finite late-time fields. The matched imported-geometry SPECTRAX-GK extraction uses the validated z_index diagnostic contract and gives normalized overlap 0.9999999994 and relative L^2 mismatch 3.33e-5 against the frozen GX raw mode.

Nonlinear Validation

Lane	Observable	Reference	Status	Baseline gate
Cyclone ITG	heat-flux window mean/std/RMS, Wphi, Wg	GX	Closed	current release gate <= 1e-1; mature target <= 5e-2 after further tightening
Cyclone Miller	same as above	GX	Closed	allow documented low-amplitude / overlap-only adjustments
KBM	Wg, Wphi, Wapar, heat flux	GX	Closed	mature lane
W7-X	heat-flux windows, saturation trend	GX + W7-X benchmark conventions	Closed	release gate <= 1e-1; manuscript target tighter where feasible. The exact-state convention audit at docs/_static/w7x_exact_state_audit.png separately closes startup state, late geometry/field arrays, and scalar diagnostic reconstruction against GX dumps with maximum finite pointwise relative error 4.62e-5 under a 1e-4 gate and scalar diagnostics below 1.8e-7.
HSX	heat-flux windows, saturation trend	GX / internal frozen references	Closed	near-threshold behavior documented
ETG full-GK pilot	short-window nonlinear transport	GX + ETG operating-point convention	Exploratory	manuscript use only if the pilot is explicitly framed as such
kinetic-electron Cyclone	electromagnetic nonlinear transport	GX	Deferred	keep out of the paper until branch identity and runtime cost are closed

Frozen artifact paths for the currently closed nonlinear lanes:

• docs/_static/nonlinear_cyclone_diag_compare_t400.png

• docs/_static/nonlinear_cyclone_miller_diag_compare_t122.png

• docs/_static/nonlinear_kbm_diag_compare_t400_stats.png

• docs/_static/nonlinear_w7x_diag_compare_t200.png

• docs/_static/hsx_nonlinear_compare_t50_true.png

• docs/_static/benchmark_core_nonlinear_atlas.png

Machine-readable nonlinear window gates are now tracked for the first refreshed subset:

The summary panel above is generated by tools/plot_nonlinear_window_statistics.py from the frozen gate-summary JSON files. It plots the gate statistic (windowed mean relative mismatch) and the maximum relative mismatch for each diagnostic, excluding exploratory summaries with gate_index_include=false.

• docs/_static/nonlinear_cyclone_miller_gate_summary.json: passed at the tightened case gate 0.095.

• docs/_static/nonlinear_kbm_gate_summary.json: passed at the tightened case gate 0.02.

• docs/_static/nonlinear_hsx_gate_summary.json: passed at the tightened case gate 0.05.

• docs/_static/nonlinear_w7x_gate_summary.json: passed at the current 0.10 mean-relative release gate after the corrected adaptive state continuation and GX-style Phi2 artifact refresh.

• docs/_static/nonlinear_cyclone_gate_summary.json: passed on the mature Cyclone t=100..400 transport window at the current 0.10 mean-relative release gate.

• docs/_static/nonlinear_cyclone_short_gate_summary.json: retained only as an exploratory t=5 startup/resolved-spectrum audit and excluded from the release-gate index.

Quasilinear Diagnostics and Model Selection

The quasilinear verification surface is deliberately split between validated linear-state diagnostics, rejected absolute-flux calibration attempts, and one scoped model-selection result. A closed model-selection status must not be read as a promoted runtime predictor.

Lane	Observable	Reference or artifact	Status	Baseline gate
Electrostatic quasilinear weights and spectra	heat/particle weights, growth/frequency spectra, and channel metadata	docs/_static/quasilinear_*_spectrum.* and docs/_static/quasilinear_validated_calibration_inputs.json	Closed as diagnostics	electrostatic channel validation and reproducible spectrum generation; this is not calibrated absolute-flux prediction
One-constant and simple saturation-rule absolute-flux models	train/holdout heat-flux prediction error	docs/_static/quasilinear_stellarator_train_holdout_report.json and docs/_static/quasilinear_saturation_rule_sweep.json	Rejected / unpromoted	current one-constant and simple-rule reports fail the held-out absolute-flux gate and must not be exposed as a user-facing saturation law
spectral_envelope_ridge model selection	leave-one-geometry-out error and interval coverage	docs/_static/quasilinear_candidate_uncertainty.json and docs/_static/quasilinear_model_selection_status.json	Closed as scoped model-selection result	the accepted candidate is a manuscript model-selection result only; the status gate does not promote a runtime/TOML absolute-flux predictor, universal nonlinear transport model, or shipped saturation option
Future absolute-flux promotion	calibrated heat-flux prediction on nonlinear holdouts	future late-window convergence metadata and promotion JSON	Open	every holdout needs finite passed post-transient convergence metadata: transient cutoff, running-mean drift, block/bootstrap uncertainty, finite sample count, and source provenance

These gates do not change the deferred W7-X lanes: W7-X zonal long-window recurrence/damping and W7-X TEM / kinetic-electron validation remain outside the current manuscript/release scope. They also do not promote a universal absolute-flux model. Production nonlinear optimization is promoted only for the selected optimized-equilibrium audit now attached to the guard; nonlinear turbulence gradients and broad multi-surface claims remain separate gates.

Autodiff Validation

Workflow	Observable	Validation type	Status
Sensitivity analysis	gamma, omega, transport windows	finite-difference / complex-step / tangent consistency	Open
Two-mode inverse problem	planted parameter recovery	gradient check + covariance estimate	Closed
UQ / Laplace example	posterior covariance and propagated uncertainty	Hessian/Jacobian validation	Open
Stellarator optimization prototype	low-dimensional objective reduction	gradient consistency + constrained solve behavior	Closed for reduced objective plumbing; open for production nonlinear heat-flux optimization

The single-mode inverse figure is intentionally a sensitivity and non-identifiability demonstration. The two-mode figure is the closed parameter-recovery validation. Both examples now write finite-difference Jacobian checks, Jacobian rank/condition number, covariance, standard deviations, correlations, and one-sigma UQ ellipse area into their summary JSON files. Those metadata are part of the validation gate: differentiated observables are not promoted to inverse-design or UQ claims unless the derivative check is conditioned and the inverse problem is identifiable.

Differentiable Geometry and Stellarator Objectives

The VMEC/Boozer objective lane is split into three claim levels. The first two are currently release/manuscript scoped; the third remains a future promotion gate.

Workflow	Observable	Reference or artifact	Status	Baseline gate
VMEC/Boozer equal-arc geometry parity	bmag, bgrad, gradpar, zero-beta metric profiles, drift profiles, q, s_hat, and solver Jacobian	docs/_static/vmec_boozer_parity_matrix.json	Closed for artifact-passing rows	mboz,nboz >= 21 and row-level parity tolerances pass; QI scope is limited to the fixed-resolution row plus evaluated robustness variants
Solver-ready objective gradients	linear eigenfrequency, growth, <k_perp^2>, electrostatic heat-flux weight, and gamma Q_i/k_perp^2	docs/_static/solver_objective_gradient_gate.json and docs/_static/vmec_boozer_gradient_holdout_matrix.json	Closed for reduced QH/Li383 gates	implicit AD/finite-difference mismatch remains within the tracked gate; current combined maximum relative mismatch is about 2.7e-2 after adding reduced nonlinear-window estimator rows
VMEC/Boozer aggregate optimization promotion	aggregate objective decrease plus surface/field-line generalization	docs/_static/vmec_boozer_aggregate_objective_gate.json, docs/_static/vmec_boozer_multi_point_objective_gate.json, docs/_static/vmec_boozer_reduced_portfolio_guard.json, docs/_static/vmec_boozer_aggregate_line_search_gate.json, docs/_static/vmec_boozer_aggregate_line_search_comparison.json, docs/_static/vmec_boozer_aggregate_alpha_holdout_gate.json, docs/_static/vmec_boozer_aggregate_surface_holdout_gate.json, docs/_static/vmec_boozer_second_equilibrium_aggregate_gate.json, docs/_static/vmec_boozer_aggregate_holdout_promotion_gate.json, and tools/check_vmec_boozer_aggregate_holdout_gate.py	Open for production transport claims	aggregate finite-difference and line-search artifacts must pass on the same training sample set, then an independent passed production-scope validation artifact must cover a held-out surface_index or field-line alpha. Production nonlinear optimization promotion additionally requires a passed replicated nonlinear-window ensemble artifact, so a single converged window cannot by itself support an optimized-equilibrium heat-flux claim. The multi-alpha finite-difference artifact passes and the growth-vs-QL comparison shows objective-dependent descent directions. The reduced-portfolio guard now verifies that the multi-alpha rows have real VMEC/Boozer provenance, multi-alpha/multi-k_y metadata, finite aggregate FD fields, finite growth/QL AD/FD diagnostics, and no production nonlinear claim. The alpha-heldout and surface-heldout splits pass as reduced generalization evidence, and Li383 passes as a second-equilibrium aggregate finite-difference/line-search check. The aggregate artifacts remain reduced optimizer-plumbing evidence. The separate production nonlinear optimization guard now includes long-window D-shaped/circular holdouts and the selected optimized-equilibrium seed/timestep audit; broader nonlinear turbulence gradients and multi-surface transport optimization are still separate gates.
Reduced stellarator ITG optimization and UQ	objective reduction history, AD/finite-difference derivative parity, local covariance, and projected uncertainty	docs/_static/stellarator_itg_optimization_comparison.json and docs/_static/stellarator_itg_optimization_uq.json	Closed as reduced optimization plumbing	objective/UQ metadata pass for the tracked QA control vector; the nonlinear entry is a smooth reduced window estimator
VMEC/Boozer nonlinear startup FD audit	compact startup-window heat-flux response to geometry perturbation	docs/_static/vmec_boozer_nonlinear_window_fd_audit.json	Exploratory plumbing gate	finite-output and finite-difference-response checks pass, but transport_average_gate = false
Selected optimized-equilibrium nonlinear transport audit	optimized-equilibrium post-transient heat-flux average with uncertainty and nonlinear audit bars	docs/_static/optimized_equilibrium_replicates/optimized_equilibrium_replicate_t700_ensemble_gate.json and docs/_static/production_nonlinear_optimization_guard.json	Closed for selected optimized-equilibrium replicated transport audit	the selected QA optimized equilibrium passes the t=[350,700] seed and timestep ensemble gate; nonlinear turbulence-gradient and broad multi-surface/multi-field-line optimization claims remain open

Use this section as the verification boundary for README figures: the VMEC/Boozer parity, gradient-holdout, and reduced optimization/UQ panels can be cited as reduced objective evidence. Startup-window finite-difference panels and reduced nonlinear-window estimators must not be cited as saturated transport-gradient validation. The optimized-equilibrium replicate panel may be cited as a post-transient transport-window audit for the selected candidate, not as a universal quasilinear absolute-flux model.

Parallelization Validation

Independent k_y and ensemble parallelization is accepted only when a serial numerical-identity gate accompanies the timing data. The current closed artifact is docs/_static/parallel_ky_scan_gate.png with metadata in docs/_static/parallel_ky_scan_gate.json. It runs the real Cyclone linear solver with ky_batch=1 and a fixed-shape batched scan, then requires max_gamma_rel_error <= 1e-8 and max_omega_abs_error <= 1e-8. The observed speedup is reported as an engineering metric, not as the gate itself. The larger CPU/GPU strong-scaling artifacts docs/_static/independent_ky_scan_scaling_large.json and docs/_static/quasilinear_uq_ensemble_scaling_large.json are the current release references for production independent-work scaling. Their split CPU and GPU companions must keep per-row identity, timing, and worker/profile metadata synchronized with the performance and validation manifests. docs/_static/parallelization_completion_status.json is the release ledger that turns those artifacts into a scoped production-closure claim while keeping nonlinear domain-decomposition speedup out of scope. Fixed-step nonlinear full-state sharding now has an engineering identity artifact at docs/_static/nonlinear_sharding_profile.json generated by tools/profile_nonlinear_sharding.py. That closes the control-flow and final-state identity layer for the pjit state-sharded scan. The corresponding two-GPU office artifact, docs/_static/nonlinear_sharding_profile_office_gpu.json, confirms active auto/kx sharding with zero final-state error on the bounded profiling grid. The large combined sweep docs/_static/nonlinear_sharding_strong_scaling_large.json is a profiler/identity artifact, not a production nonlinear speedup claim. The same diagnostic keeps z-axis FFT sharding out of the release claim until it has a dedicated communication/layout design and a passing identity gate. True nonlinear domain decomposition with halo/FFT communication, conservation checks, and benchmark-size speedup remains outside the release claim.

Notes

• A lane should not move from Open to Closed without an owning script, frozen artifact path, and literature/reference statement.

• README figures should use only Closed lanes unless a panel is explicitly marked exploratory.

• Raw eigenfunction overlays for manuscript use should be rendered only from frozen reference bundles checked into docs/_static/reference_modes/. Do not build publication figures from transient external files or ad hoc office-machine outputs.

• Experimental-facing figures such as W7-X fluctuation spectra should remain scoped as simulation diagnostics unless the diagnostic transfer function and access model are encoded directly in the repo.

• Electromagnetic stellarator claims should be split explicitly into heavy-electron verification and realistic-electron research runs.