# HBF implementation requirements and research evidence review

Prepared 11 September 2026. The full supplied 1,114-line plan, evidence README, code findings, numerical records and source inventories were read. This document specifies what software can credibly deliver now and what remains a research, data-access or institutional dependency. It does not certify the parent implementation, a deployed release, an experiment or a reactor. Implementation acceptance must be attached by the integrating agent to the requirement IDs below.

The central decision remains: make one charged-particle measurement and conversion workflow independently inspectable. Repair correctness first; connect the existing particle/heating/collector tools; implement calibrated data registration and reproducible evaluation; then obtain the measurements needed to assess physical predictions. An interface can be completed now while its experimental endpoint remains pending.

This review uses task-specific public primary sources. It did not obtain private industrial/university records, contact prospective partners, apply for grants, execute external commercial solvers or survey all software worldwide. Documentation access is distinct from downloading, licensing, integrating and experimentally validating a dataset. The source register records that distinction per item.

## Evidence already supplied

The 9 September evidence is historical, including a CPU-compatible browser path. It should remain unchanged and retain its date and hashes. Source captures cover selected public files, not the entire development repository. The benchmark record identifies appVersion 7.4.7 and kernel 1.0.0 while the historical UI release was 7.4.9; that version separation must be preserved, not normalized to one invented release.

- Fourteen of fourteen supplied Node numerical fixtures were recorded as passing on Node 24.19.0/Linux x64. This research read those records; it did not freshly rerun them. The integrating agent should report its own new run separately.
- Four Python/SciPy reference arrays were recorded as exactly reproduced with NumPy 2.3.5 / SciPy 1.17.0. That was a rerun of the supplied independently implemented reference, not a newly authored third solver.
- The baseline fusion/radiation ratio near 1.518 coexists with negative included-channel heating. The default reactor heating closure at η = 0.5 adds approximately 1,974.735 MW required wall heating and yields roughly −1,719.95 MW adjusted net under the model assumptions.
- Synthetic AlphaTrack F1 values near 0.94 / 0.955 concern generated shapes. They establish no real-plate species, energy or cross-site accuracy.
- DirectConvert's ideal ledger and planar current screen are already present. Energy conservation of that model is necessary numerical evidence and does not establish an apparatus efficiency.
- The historical live-browser export link was visible, but downloaded-file integrity was not established; local package round-trip success is a different observation.

## B01–B09 repair matrix

Every row is software work available now. Completion requires verification against the actual integrated build. Historical bug evidence is not evidence that a new implementation remains broken, nor evidence that a patch is complete.


| ID | Priority | Required outcome | Acceptance |
| --- | --- | --- | --- |
| B01 | P1 | Installation checker expects the wrong release | the deployed matching release reaches every configured asset check; a deliberately older/newer manifest fails clearly; an HTML error page with HTTP 200 and an incorrect MIME type still fails. Test both domain-root and subfolder deployment. |
| B02 | P1 | Exactly 200 keV is incorrectly outside the stated benchmark range | exact equal-temperature endpoints pass; 199.999 and 500.001 fail; unequal-temperature cases and all lab consumers agree; table, badge, JSON, and CSV classifications match. |
| B03 | P2 | Research provenance reports a different release from the running app | a package generated by 7.4.9 records that app release, retains the correct model/data versions, and round-trips through every migration explicitly advertised. Preserve historical packages unchanged. |
| B04 | P2 | Manual calculation can leave the live helper saying “Recalculating…” | rapid edit→Run, edit→Reset, invalid→valid, and edit during an outstanding job produce mutually consistent messages and no stale result advertised as current. |
| B05 | P2 | A/B results omit units for many scientific quantities | all primary and changed A/B quantities have an explicit unit, definition, and valid-null meaning; number formatting is derived from the quantity rather than the internal property name. |
| B06 | P3 | The comparator overstates meaningful change | changing density shows the expected density, rate, power, and inventory effects clearly; invariant ratios are marked unchanged within numeric tolerance; the raw view preserves every difference. |
| B07 | P2 | Collector stage labels overlap in 3D | every stage can be identified at the default desktop view and narrow widths without overlapping labels; the table and view highlight the same stage. |
| B08 | P2 | Default all-panel scenes hide the object being explained | first view shows the relevant apparatus and the primary result together; opening a panel recalculates available camera bounds; focus remains visible while a panel scrolls. |
| B09 | P2 | Important research tools require editing raw JSON | a user can set heating efficiency, configure a threshold study, and import a particle-flux table without hand-editing JSON; the exported record remains equivalent to the advanced representation. |


B01 must derive the expected release from an independent build-owned identity and compare page, boot, checker and manifests. Merely reading an arbitrary server manifest as the expectation cannot detect a mutually consistent wrong release. Keep deliberate mixed-release, wrong MIME and HTTP 200 HTML-error fixtures. Test root and subfolder routes, and record the complete check inventory.

B02 requires a single documented range function across UI/badge/JSON/CSV and scientific consumers. For equal temperatures return the exact entered identity; for unequal temperatures use a stable weighted combination. An allowance of 8 machine epsilons scaled to endpoint magnitude is a numerical classification tolerance, not a physical uncertainty interval. Preserve unrounded exported values and reject finite-domain failures. Include both equal endpoints, 199.999, 500.001 and asymmetric temperature pairs.

B03 must distinguish app release, kernel, schema, dataset, source/build hashes and supported migration versions. Do not rewrite historical records. B04 must use job revision/state ownership; a finally block cannot blindly replace an invalid/failed status with complete. B05/B06 must share the typed quantity registry, separate meaningful scientific comparison from raw byte/numeric differences, and use quantity-specific absolute/relative comparison tolerances unrelated to physical error bars. B07/B08 require observed keyboard and narrow-width interaction checks; CSS edits alone cannot establish those acceptance criteria. B09 must expose the parent run and make connection of an independent study deliberate.

## All twenty ranked additions and feasible completion boundaries

The original effort estimates remain person-weeks of focused work and can overlap. They exclude elapsed delays for data agreements, facilities, ethics/safety review where applicable, procurement and scientific supervision. These additions must extend existing tools; Poincaré plots, G-EQDSK import, Monte Carlo, power ledgers, segmentation and research exports are already present in the audited platform.


| Rank/ID | Addition | Deliverable possible now | Required gate beyond software |
| --- | --- | --- | --- |
| 1 / F01 | Consistent release metadata, checker, and run state | Build-owned release configuration, manifests, exports and job state | No external data needed for regression; deployment checks need actual hosted release |
| 2 / F02 | Connected heating/electricity account | Compose heating and terminal electricity into main scenario | Numeric ledger only until components have measured evidence |
| 3 / F03 | Calibrated plate registry and locked evaluation split | Immutable plate/calibration registry and locked split tooling | Real calibrated images, rights and independent plate/session/site labels required |
| 4 / F04 | Robust track segmentation with uncertainty and abstention | Keep baseline and implement evaluation/abstention interface | Learned model training, calibration coverage and cross-site performance require labels and held-out plates |
| 5 / F05 | Joint particle-flux transfer between labs | Connect existing particle flux format to all consumers with units/frame/rate checks | Device-source authenticity requires measured/reviewed inputs |
| 6 / F06 | Measured collector I–V and power comparison | CSV I–V import, precommit prediction and residual tools | Non-fusion facility, electrical/source calibration and held-out sweeps required |
| 7 / F07 | Correlated uncertainty and model-discrepancy propagation | Joint realizations and correlated nuisance variables, explicit discrepancy | Covariances and coverage claims require defensible data; unknown omissions remain unresolved |
| 8 / F08 | Independent replay and reference adapters | One explicit solver adapter, complete input/output package | External runtime/licence/reference operator and matched problem needed for independent execution |
| 9 / F09 | Measured field-map and diagnostic residual views | Field-map coordinate/calibration import and withheld-point residuals | Measured geometry/field and independent diagnostics needed |
| 10 / F10 | Detector response/inverse spectrum with identifiability checks | Response matrix import, Poisson model and identifiability diagnostics | Known response exposures; fail to unknown when non-identifiable |
| 11 / F11 | Next-experiment selection by uncertainty reduction and cost | Candidate/cost registry and transparent design baseline | Validated surrogate plus repeated locked campaigns needed for benefit claims |
| 12 / F12 | Source covariance, differential data and model comparisons | Data parsing, covariance shape/PSD checks, model comparisons | Actual evaluated covariance/differential records and redistribution review needed |
| 13 / F13 | Batch images, tiled processing and annotation adjudication | Batch manifest, overlap tiling, annotation disagreement/adjudication | Runtime and boundary reconciliation must be verified with partner-sized images |
| 14 / F14 | Explicit evidence status per component and claim | Observable-specific V0–V5 evidence records and dependency links | Only real accepted evidence can advance a claim |
| 15 / F15 | Time-dependent species/ash and sustainment model | Scoped reduced species/ash ODE prototype with explicit conservation | Plasma supervisor, maintained distribution/collision and extraction assumptions needed |
| 16 / F16 | Collector field/particle coupling with a reference solver | Reference-job configuration and comparison interface | Low-current measurements first, then self-consistent coupling justified by regime/residuals |
| 17 / F17 | Dimensionless scaling from demonstrator to target regime | Dimensionless scaling worksheet and regime mismatch warnings | Similarity is necessary context, not full reactor-scale validation |
| 18 / F18 | Multi-institution review, data permissions, and embargo support | Local review package and explicit permissions metadata now | Hosted authorization/embargo enforcement needs backend and actual institutional agreement |
| 19 / F19 | Free-boundary equilibrium or advanced stability adapters | Document existing fixed-boundary scope; adapter only for selected experiment | Do not pretend a JSON import or field-line plot establishes free-boundary/stability solver |
| 20 / F20 | Additional visual polish and guided teaching modules | Focused accessible task flow and teaching improvements | Visual polish does not advance physical evidence level |


A practical release can supply forms, strict import contracts, baseline algorithms, synthetic fixtures, replay tools, uncertainty calculations, claim review and proposal templates. It cannot honestly mark F03/F04/F06/F09/F10/F11/F15/F16/F18/F19 complete in their scientific or institutional sense solely because corresponding controls exist. A documented planned interface is also distinct from an executed adapter.

## Contracts and implementation acceptance

The draft JSON contains complete tests and scope for every contract below. These are semantic requirements; they do not require thirty packages or separate services. Keep a small deterministic core and extend the existing scientific record format.


| ID | Contract | Required content | Acceptance boundary |
| --- | --- | --- | --- |
| C01 | Deterministic core and incremental modules | Separate numerical core from DOM/rendering; No silent source fetch in scientific kernel; Use workers for long jobs; retain browser/local-file first architecture; No forced nine-service migration | Same canonical scientific inputs yield reproducible scoped output; Rendering changes do not change scientific input hash |
| C02 | Versioned RunManifest | schemaVersion/runId/appRelease/kernelVersions/sourceCommit/artifactHashes/inputHash/parentRunIds/createdAtUtc/evidenceClass; datasets/assumptions/coordinateFrames/randomStreams/numericalSettings/validation; Define migrations; preserve historical packages | Complete dependency closure and raw scientific inputs recorded; Unknown versions remain explicit null with reason, never invented |
| C03 | Dataset and rights identity | id/version/rawSha256/normalizedSha256/citation/sourceUrl/retrievedAtUtc/permissionRecordId/transformRecordId; Retain raw bytes and create new versions on correction; Original units/frame/observable and licence owner recorded | No default source changes without result-impact review; Public derivative contains only permitted materials |
| C04 | Assumptions and validation records | Assumption definition/value/unit/evidence/status/affected outputs; Validation quantity/reference/tolerance/result/scope/reviewer/date | Claim traced to individual observable and exact supporting evidence |
| C05 | Null semantics and units | Explicit quantity registry and null reason; Nonfinite numbers rejected or explicitly unavailable; no silent conversion to zero | Every displayed/exported quantity has unit, definition and null meaning |
| C06 | Scientific hashing | Hash scientific inputs separately from timestamp/camera/view; Record transitive imports and data transforms; Hash does not authenticate source | Cosmetic changes preserve input hash; Scientific input/dependency mutation changes recorded identity |
| C07 | Joint particle transfer | Extend hbf-particle-flux@1; species/charge/mass; surface/normal/frame; paired position/direction or SI momentum; energy kinetic convention; time; weight kind; total rate; source parentage; fate; Rotation/boost/transformation provenance; No hidden renormalization after losses | Known rotations/boosts and inverse transforms agree; Fraction weights sum to one at declared boundary and rate carried separately; Lost rate and energy preserved in fate ledger; Reject missing conventions, negative weights, inconsistent energy/momentum |
| C08 | Energy origin graph | Unique origins, sources/targets, boundary and time interval; Units and uncertainty/evidence status on streams; Fusion deposition and converter allocation partition a common source; Grid/DC/component auxiliary boundaries explicit | Node energy balance residual retained; No terminal double-credit of electricity/heat or plasma-deposited alpha energy |
| C09 | Heating closure | Plasma loss minus deposited alpha gives absorbed deficit; Wall-plug efficiency strictly >0 and <=1; Subtract only additional heating absent from auxiliaries; Driver waste heat separate origin and temperature | Primary scenario shows conditional and heating-adjusted values and all assumptions; No adjustment implies measured plant result |
| C10 | Heat reservoirs | Per-origin temperature and capture; Recovery efficiency model and Carnot/domain checks; Unknown temperature means no unsupported common-hot credit | Ledger energy closes within typed numerical tolerance; Unaccounted measured residual kept visible with uncertainty |
| C11 | Calibrated image decoding | Decode once, orientation/colour handling/pixel hash; Physical pixel scale with uncertainty and calibration ID; TIFF/OME only if required; analysis resolution distinct from preview | Uncalibrated image may produce pixel morphology but not false physical size/species |
| C12 | Tiling and batch identity | Overlap exceeds supported object radius; Global physical-coordinate deduplication; Exactly-once object assignment and edge policy | Tiled and untiled results agree on overlap-capable fixtures; Limit error offers actionable tiling route |
| C13 | Annotations and frozen split | Retain all annotators, disagreement and adjudication; Split by plate/session/site; Development, tuning and locked evaluation subsets separate | No same-plate crop leakage; Evaluation lock/method hash predate evaluation opening |
| C14 | Segmentation and uncertainty | Preserve Otsu/components baseline; Compare watershed/instance and learned baseline only with labels; Miss/false/split/merge/count/size/time metrics; Abstain for missing calibration, severe overlap, domain shift | Eligible-class count bias target preregistered; Coverage tested on independent plates; no measured-performance claim from synthetic circles |
| C15 | Detector inverse response | Response matrix R(theta), background b and nonnegative source phi; Likelihood suited to counting unit; Residuals, identifiability, regularization/prior sensitivity | Independent response calibration and exposures; Unknown/interval/multiple solutions when non-identifiable |
| C16 | Mass covariance | Net stoichiometric coefficient vector with alpha coefficient -3; Actual covariance mapping and atomic/nuclear conventions; Convert u-to-MeV consistently | PSD/symmetry/dimension checks; Repeated nuclide treated as one correlated evaluated quantity; Unknown covariance not fabricated |
| C17 | Cross-section covariance | Statistical and common normalization uncertainties separate; 3.4% normalization drawn once per realization; Unknown extrapolation/covariance disclosed; no independent-per-row fiction | Reactivity uncertainty propagates full available covariance; Numerical refinement and physical/source uncertainty reported separately |
| C18 | Correlated sampling | Joint state once per realization through all modules; Shared calibration not independently resampled; Purposeful random stream algorithm/seed; Unknown omitted mechanisms retained unresolved | Positive/negative correlations preserved in fixture; No ordinary independent-input Sobol interpretation for arbitrary correlated inputs |
| C19 | Surrogate validity | Verified underlying solver and domain; Independent/boundary/failure test cases; Paired accuracy and speed measures | Out-of-domain predictions abstain; No uncertainty calibration claim before coverage evaluation |
| C20 | Job state and history | idle->queued->running->complete/invalid/failed/cancelled; One input revision and active job; Edits make earlier result historical; Manual run cancels debounce; cleanup does not overwrite errors; Progress reflects completed work; playback time separate | Outdated/cancelled jobs cannot replace current result; Historical run can be inspected/exported with original inputs |
| C21 | Typed studies and imports | Forms for heating, threshold study, particle table mapping; Units/examples/schema error locations/parent-run preview; JSON advanced representation preserved | Typed and JSON settings export equivalent scientific record; User explicitly connects a separate study to parent scenario |
| C22 | External reference job | Choose one matched solver family; Complete geometry/inputs/BC/coordinates/equations/numerics exported; Solver version/input/data hashes and environment recorded; Import observables/diagnostics, not screenshots | Executed reference results distinct from mocked fixture; Mismatch in physics/boundaries blocks equivalence claim |
| C23 | Collaboration service | Only after need: project permissions/immutable storage/job status/resource limits/audit; Restricted files remain agreed environment; Comments attach fixed run; changed assumptions create new run | Local metadata never represented as enforced server authorization; Two institutions and rights agreement before service claims |
| C24 | Reproduction package | README/manifest/permitted inputs/outputs/calibration/environment lock/methods/limits/executable entry; RO-Crate mapping with pinned spec; scientific schema retained | New operator replays selected output; integrity/migration failures actionable; Run package completion distinguished from browser download verification |
| C25 | Measured collector experiment | Incident rate/energy/direction joint source and geometry; Applied and measured potentials/current signs; Offsets/leakage/background/source-on/off/timestamp/calibrations/load boundary; Freeze prediction before holdout; second setting/session without retune | Residuals assessed against preregistered measured/model uncertainty; Loss terms improve independent observable; unknown energy unaccounted |
| C26 | Dimensionless scaling | E/q; qV/E; angular spread; gyroradius/size; collisionality; perveance/current; surface regime; Keep low-current geometry validation separate from MeV/current/material scaling | Regime differences block unjustified target-regime extrapolation |
| C27 | Reduced maintained nonthermal/ash model | Distribution normalization and required sustaining power; Slowing/scattering/radiation/loss terms; Species inventory/helium dilution/extraction-heating tradeoff; No arbitrary gain multiplier | Conservation and known limits first; Expert reviewed domain before experimental claims |
| C28 | Secondary channels and materials | Explicit reaction/material/energy coverage; Appropriate external transport data; No zero-radiation or zero-activation badge from aneutronic label | Specialist scope and source licence/processing identity documented |
| C29 | Rights-aware ingestion sequence | Register->retain raw->parse->normalize->structural validation->scientific suitability->permitted derivative->deliberate updates; Finite/monotonic/duplicate/dimension/normalization/channel/covariance checks; Preserve estimated/unknown flags and energies with shifts | Parser pass cannot advance measured authenticity; Unknown or restricted permission blocks bundling, not silently auto-approved |
| C30 | Novelty and AI boundaries | Focused prior-art/baseline comparison per precise method; No worldwide novelty claim from bounded search; No chatbot certification, fabricated data, silent scientific changes or equipment actuation | AI evidence records training provenance, frozen methods, data split and failure modes; Automation limited to approved acquisition planning/records |


## Critical equations, domains and independent checks

The following equations are implementation requirements and reproducible reference targets. Their validity is conditional on the stated model. Derived accounting identities are not measured predictions.

### Reactivity and source power

Use temperatures as energy parameters θ=kBT, in joules inside the SI integral:

\[
\theta_{rel}=\frac{m_B\theta_p+m_p\theta_B}{m_p+m_B},\qquad
\langle\sigma v\rangle=\sqrt{\frac{8}{\pi\mu}}\theta_{rel}^{-3/2}\int_0^\infty E\sigma(E)e^{-E/\theta_{rel}}\,dE,
\quad \mu=\frac{m_pm_B}{m_p+m_B}.
\]

For distinguishable p and B reactants, reaction rate density is np nB〈σv〉 and fusion power density is that rate multiplied by reaction energy Q. The supplied constant-cross-section analytic fixture yields σ0√(8θ/(πμ)); it independently tests integration and conversion factors. It does not validate the finite published p–B table, missing tails, differential branches or a maintained nonthermal plasma.

One 3.4% common cross-section normalization factor is one random variable across the whole curve. In a quadrature representation with weight vector a, linearized source variance is aᵀCσa; this is only as valid as the supplied covariance. Do not replace undocumented correlation by independent row error bars. Keep numerical quadrature error, source evaluation difference and model discrepancy distinct. The retained D01 table is an HBF-derived 41-point evaluation, not a new experimental dataset.

### Reaction energies and repeated species

For net mass coefficients a, Q=c²aᵀm and Var(Q)=c⁴aᵀCma. If masses are u, use the appropriate u-to-MeV constant consistently. The three outgoing alphas have coefficient −3 for the same evaluated mass: their uncertainty is not three independent draws. Atomic masses, nuclear masses and electron binding differ at precision relevant to a quoted uncertainty. AME2020 lists covariance data, but this review could not retrieve that ZIP or validate the needed indexing. [AMDC](https://www-nds.iaea.org/amdc/), [CODATA2022](https://physics.nist.gov/cuu/Constants/Table/allascii.txt).

### Particle transfer and collector ledger

A flux boundary with normalized sample weights has Σwi=1, incident rate Ṅ and input power Pin=ṄΣwiEi. A selected population has current I=qṄΣselectedwi for one declared species/charge, with sign conventions explicit. Positive-ion retarding work qΔV is valid only for the specified trajectory/potential path; aperture loss, angular energy, reflection, secondary electrons and charge exchange require their own model and measured assumptions. When a fraction is lost, retain its original weight/rate and energy in a fate account. Renormalizing only survivors while retaining the original total rate creates artificial power.

For a simple prescribed planar test, incident parallel kinetic energy Eparallel=Ecos²θ defines the ideal retarding threshold. This is a verification fixture, not a general collector solution. Self-consistent charged-particle models couple ∇²φ=−ρ/ε0 with dp/dt=q(E+v×B); adding a space-charge warning does not solve that coupled problem. SIMION's documentation explicitly distinguishes approximate repulsion estimates from a Poisson treatment. [SIMION](https://simion.com/info/space_charge.html), [WarpX parameters](https://warpx.readthedocs.io/en/latest/usage/parameters.html).

### Heating and energy origin

\[
P_{heat,wall}=\max(0,P_{plasma,loss}-P_{\alpha,dep})/\eta_{wall\to absorbed},
\quad P_{net,grid}=P_{electric,gross}-P_{heat,wall}-\sum P_{other,grid}.
\]

Require 0 < η ≤ 1. If a baseline auxiliary already includes heating, subtract only the additional amount needed; reject incompatible boundaries instead of charging it twice. Each energy node obeys Ein+Egeneration−Eout−ΔEstored=Eunaccounted. Keep measured Eunaccounted with uncertainty, even when it is nonzero. Fusion alpha deposition and converter input are partitions, and collector electricity plus residual heat are partitions. External electrical heating and its waste heat retain their external energy origin. A thermal recovery coefficient cannot exceed its justified limit for stated hot/cold temperatures; unknown temperatures cannot be silently replaced with a common high temperature.

The default negative adjusted net is an instructive historical case. It demonstrates why conditional output alone is misleading, not why a real apparatus has been measured to produce that number. [Ochs–Fisch](https://arxiv.org/abs/2310.18508), [adiabatic conversion theory](https://arxiv.org/abs/2410.15485).

### Fields and orbits

With ψ as poloidal flux per radian and B=∇ψ×∇φ+F∇φ, BR=−ψZ/R, BZ=ψR/R and Bφ=F/R. The axisymmetric equilibrium operator is Δ*ψ=ψRR−ψR/R+ψZZ with source−μ0R²p′(ψ)−FF′(ψ). Coordinate and COCOS conventions must be explicit at import/export. A sign flip or 2π convention error can retain visually plausible field lines while invalidating comparisons.

Use analytic uniform-B gyromotion, current-loop axis field, flux preservation, event localization and grid/timestep refinement as independent numerical checks. Full orbit, guiding centre, field-line arclength and collisional transport are different models. A one-microsecond orbit without a wall hit establishes survival only over that interval. Manufactured GS residual and cross-code agreement assess the stated discrete/fixed-boundary problem, not physical confinement or free-boundary stability. The external MAST-U study is a useful matched-reference precedent but uses reconstructed EFIT++ inputs; it is not an exact ground-truth equilibrium. [ASCOT5](https://ascot4fusion.github.io/ascot5/), [FreeGSNKE/Fiesta study](https://arxiv.org/html/2407.12432v5).

### Radiation

The NRL hydrogen-like baseline is PBr=1.69×10⁻³² ne sqrt(Te)ΣZ²ni W/cm³ when ne and ni are cm⁻³ and Te is eV. Thus the coefficient with SI densities and output W/m³ is 1.69×10⁻³⁸. This provides the supplied NUM-22 unit oracle. Relativistic and electron-electron contributions need a model with a declared domain; a high-temperature calculation must not inherit the low-temperature baseline's qualification. Xie's public industry-affiliated research gives fitting/reference comparisons with its own domains; no private plasma records were accessed. [NRL2023 Eq 30, page 58](https://www.nrl.navy.mil/Portals/38/PDF%20Files/NRL_Plasma_Formulary_2023.pdf), [Xie2024](https://arxiv.org/html/2404.11540v2).

### Detector response and measured residuals

A possible count-like model is y∼Poisson(R(θ)φ+b), with detector/calibration nuisance parameters θ and background b. This is not automatically the correct likelihood for every extracted shape. Fit only nonnegative identifiable source combinations; inspect residuals, sensitivity to priors/regularization and response-matrix rank/conditioning. If different species/energy spectra fit equally well, return unknown, multiple explanations or appropriately wide intervals.

For measured predictions with shared nuisance parameters, compare residual vector r using its justified covariance C: a quadratic score rᵀC⁻¹r is meaningful only if C, effective rank and sampling assumptions are defensible. Independent pointwise error bars cannot represent a common voltage gain, pixel scale or source normalization uncertainty. Uncertainty intervals estimated from one synthetic image or a threshold sweep are not calibrated physical prediction intervals.

## Data acquisition and rights decisions

The following catalogue covers all D01–D18 items in the original plan. The software may supply parsers and schema checks now. A source becomes a supported scientific input only after its raw bytes, frame/observable, transform history, uncertainty semantics and permitted use are reviewed.


| ID | Source | Access/action | Boundary |
| --- | --- | --- | --- |
| D01 | Supplied Sikora–Weller derived CSV | Already available locally and hashed; retain baseline | Preserve derivation/source assumptions; not raw events |
| D02 | EXFOR | Landing only; selected reaction records pending | No claimed full/differential extraction or reuse licence |
| D03 | AME2020 masses | Official ASCII links verified; parser feasible | Pin values, states, conventions and paper citation |
| D04 | AME2020 covariance | ZIP listing verified; download failed here | No covariance contents/index coverage or integration claimed |
| D05 | NUBASE2020 | Official file listing verified | Retain evaluated versus estimated entries and states |
| D06 | LiveChart | CSV API documentation retrieved | Actual B-11 response/CORS not executed; preserve blanks, shifts and state energies |
| D07 | CIAAW2024 | Table/current label read | Natural abundance not enriched target assay; item rights review |
| D08 | NIST STAR/SRD124 | Reference landing read; no material tables acquired | SRD may be copyrighted; stopping power is not CR-39 etch response |
| D09 | IBANDL | Redirected landing; no record content | Exact channel/angle coverage and rights pending |
| D10 | FENDL3.2c | Official release/process note read | Specialist reaction/material/energy scope; pin raw and processed hashes |
| D11 | OPEN-ADAS | Terms read; no dataset acquired | Written permission before commercial/site/database/code redistribution |
| D12 | MAST-U equilibrium comparison | Public paper and cases identified | Repository/data licence and actual ingestion/execution pending |
| D13 | IAEA Fusion Data Lake | 2026 project report read | Federation catalogue is not blanket underlying-data access |
| D14 | Published CR-39 learned model | Paper inspected as prior work | Raw images/annotation/model licence not established; no HBF result |
| D15 | Xie radiation | CC-BY4.0 paper read, linked code identified | Code licence not inspected; public theory only, no private experiments |
| D16 | Vendor calibration | Not acquired | Actual serial/certificate/gain/drift/use permission required |
| D17 | Partner raw campaigns | Not acquired | Host/facility, raw records, calibration, rights/publication agreement needed |
| D18 | Material characterization | Not acquired | Preparation-specific measurements and permitted use needed |


The important rights distinction is concrete. OPEN-ADAS requires prior written permission for the relevant incorporation and redistribution uses; an import button is not permission. NIST's policy distinguishes copyrighted Standard Reference Data from other employee-created works, so “US government” is not a sufficient redistribution licence for SRD124. A CC-BY paper licence also does not automatically license its author's raw images, code, weights or partner material. [OPEN-ADAS terms](https://open.adas.ac.uk/terms-and-conditions), [NIST policy](https://www.nist.gov/open/copyright-fair-use-and-licensing-statements-srd-data-software-and-technical-series-publications).

FENDL documents a tungsten neutron-file synchronization correction on 31 May 2025 while processed ACE/application files were unaffected. This reinforces the need for exact files and processing hashes rather than a vague latest label. [FENDL release note](https://www-nds.iaea.org/fendl/).

Recommended source pipeline: register owner/citation/URL/version/rights/observable; retain raw bytes and hashes; parse positive and negative examples; normalize with recorded transformations; check finite values, monotonicity, duplicates, dimensions, normalization and covariance; review scientific suitability; publish only permitted derivatives; update defaults after quantified impact review. All existing runs must retain the earlier dataset version. No complete internet corpus is needed or claimed.

## Reusable external comparisons

These are selected reference opportunities, not a declaration that all listed tools were installed, their output accepted, or their code licensed for bundling.


| Candidate | Matched use | First defensible benchmark | Access/licence/physical limitation |
| --- | --- | --- | --- |
| Supplied NUM fixtures | Core regression | Analytic fields/orbits/GS/integral/unit/package cases | Historical14/14; rerun separate; V1/V2 only |
| ASCOT5 | Orbit and collisional transport adapter | Same prescribed field, species, initial conditions, event/time convention; energy/phase/loss comparison | Docs read, licence fetch failed, no execution; reduced model must match physics |
| WarpX | Electrostatic collector reference | One geometry and low-current limit, then grid/particle/space-charge refinement | Docs read, licence fetch failed, no run; boundary/species/charge model explicit |
| COMSOL/SIMION | Commercial field/trajectory comparison | Predicted I(V)/trajectory/energy at same apertures and electrodes | Vendor capability docs only; licence and operator needed |
| FreeGSNKE/Fiesta MAST-U | Equilibrium interchange/reference | One pinned 45425 or 45292 time slice with matching currents/profiles/convention | Paper retrieved, data not ingested; reconstruction uncertainty remains |
| NRL/Xie radiation | Equation and unit reference | NRL unit conversion plus Xie domain-specific e–i/e–e comparison | Public equations; no physical plasma validation |
| NIST STAR | Stopping/range cross-check | One permitted material/energy table and interpolation check | Specific SRD rights and material match required |
| Geant4 | Material/detector transport | Known slab/material/source with uncertainty and declared physics list | Cannot replace empirical etching/microscopy response |
| OpenMC | Secondary neutron/photon study | Explicit source/reaction/material benchmark after specialist selection | Not alpha collection; transport library and physics suitability separate |
| Otsu/watershed/ilastik/Cellpose | Segmentation baseline comparison | Same independent plates, fixed annotation protocol, scientific endpoint and correction time | No fabricated accuracy; code/model/data licences reviewed separately |
| PROCESS/bluemira | Integrated accounting/interface comparison | One matched component energy boundary, not whole-plant output ranking | Published general capabilities do not validate HBF p–B assumptions |
| openPMD/IMAS/RO-Crate | Interchange and packaging | Small semantic round-trip preserving units, frames, weights and provenance | Standards mapping test required; do not invent format equivalence |


External reference acceptance requires a configuration fixed independently of HBF outputs, complete solver/runtime/input identity, matched equations and boundary conditions, dimensionful tolerances, raw observable arrays and residuals. Reproducing the same implementation under a second label is not independent validation. For RO-Crate, pin the 1.2 mapping if implemented; its specification is Apache 2.0 and JSON-LD context/examples CC0. [RO-Crate1.2](https://www.researchobject.org/ro-crate/specification/1.2/), [openPMD](https://openpmd-api.readthedocs.io/en/latest/), [IMAS-Python](https://imas-python.readthedocs.io/en/latest/).

## Experiment and AI gates

### Flagship A: calibrated AlphaTrack

The immediate product is a plate registry and honest benchmark runner. Register immutable plate/image identity, original and decoded-pixel hashes, scale/uncertainty, microscope settings, acquisition session/site, detector batch, exposure/etch history and controls. Keep every annotator's masks/centroids, disagreements and adjudications. Split by plate, acquisition session and site before selecting a model; crops from one plate must not straddle train and evaluation.

Compare the existing baseline to a touching-object/instance method and a learned baseline only after suitable labels exist. Freeze preprocessing, model/weights, hyperparameters and evaluation hashes before holdout access. Report false/missed objects, split/merge errors, count/density/size bias, correction time and memory/time by acquisition condition. Include missing calibration, severe overlap, image shift, unknown format, invalid scale and out-of-distribution refusal paths. Record the sampling unit: thousands of tracks on one plate are not thousands of independent plates.

The plan's 30% manual-correction-time reduction is a proposed partner target, not an observed achievement. Set the primary count/size endpoint and noninferiority/error tolerance prospectively. Evaluate interval coverage on independent plates with uncertainty around the coverage estimate. If an image cannot support physically independent species/energy labels, publish morphometry and retain unknown. Never train on pseudo-labels generated by the unvalidated physical model being tested. Published D–D CR-39 classification provides prior art, not permission to claim its accuracy for HBF. [Wang etal.](https://arxiv.org/abs/2503.06932).

### Flagship B: measured non-fusion collector

Prepare software for an institution-operated experiment: incident/source-estimated particle rate and joint energy/angle distribution; geometry/surfaces; commanded and measured potentials; signed current readings; dark/leakage/source-off controls; timestamps and synchronization; calibration IDs and uncertainties; and the exact load/electrical measurement boundary. This document supplies no apparatus construction or operating procedure.

Sequence the evidence: analytic monoenergetic reference; preserved measured sweep; prediction preregistered before evaluation; one measured spread/divergence/aperture/background term at a time; matched external field/trajectory comparison; self-consistent space charge only when regime/residuals demand it; then a second session/configuration without retuning. Model and measurement uncertainty must explain residuals on holdout, or the model fails within that scope. Unobserved losses remain explicit. Every added mechanism should improve an independent observable.

Compare E/q, qV/E, angle, gyroradius/geometry, collisionality, perveance/current and surface state between demonstrator and target. A low-current test may validate measurement and geometry while leaving reactor current, MeV material interactions and integrated net power entirely unresolved. A reproducible negative result is still useful research.

### Flagship C: uncertainty and next-experiment selection

Choose one calibrated chain. Draw one joint nuisance state per realization and carry it through source, transport, detector and electricity. Separate counting variability, calibration, numerical error, evaluated source uncertainty, model discrepancy and unknown omissions. Never encode a wholly unknown mechanism as a narrow made-up prior.

Use transparent sensitivity/design baselines before a surrogate. A surrogate needs independent boundary and failure-region tests plus measured speed/error; outside its domain it must abstain. Experiment ranking should expose the observable, expected uncertainty reduction, cost, feasibility and constraints. Compare against random and simple space-filling selection over repeated locked campaigns before claiming a benefit. No automated laboratory control is part of this initial scope. LLM assistance may draft from structured records; it cannot certify a scientific result or silently alter inputs.

### Independent validation ladder and release gates


| Level | Evidence | Permitted interpretation |
| --- | --- | --- |
| V0 | Unchecked conceptual assumption | A scenario or hypothesis |
| V1 | Analytic/manufactured/numerical verification | Correct implementation for the tested equations and fixtures |
| V2 | Independent implementation/reference agreement | Reduced risk of implementation error within the matched specification |
| V3 | Calibrated experimental comparison | Prediction tested for the specified apparatus and operating domain |
| V4 | Independent institution/repeated campaign | Evidence of reproducibility and transfer within the tested conditions |
| V5 | Integrated system evidence under declared boundaries | Only the actually measured integrated claim; not automatic general reactor qualification |


These are proposed internal claim labels, not an external certification scheme. Assign them per observable; a verified orbit algorithm does not confer V3 on its predicted wall-loss fraction.


| Gate | Acceptance | Status required |
| --- | --- | --- |
| Release consistency | Page, boot bundle, checker, manifests and new export metadata identify the intended app release; pinned kernel versions remain explicit | Attach actual integrated evidence; pending in this research |
| Input/domain edges | Inclusive/exclusive endpoints, invalid numbers and boundary-adjacent floats match a documented contract | Attach actual integrated evidence; pending in this research |
| State/cancellation | Manual run and debounce paths end in consistent states; outdated jobs cannot overwrite current input revisions | Attach actual integrated evidence; pending in this research |
| Unit/frame contract | Every public quantity has unit/meaning; incompatible frames or weights are rejected | Attach actual integrated evidence; pending in this research |
| Numerical reference | Analytic fixtures, refinement order, conservation and event-localization criteria pass on the supported runtime set | Attach actual integrated evidence; pending in this research |
| Package reproduction | Exported scientific inputs/data references reopen correctly; corrupt/missing/wrong-version inputs fail clearly | Attach actual integrated evidence; pending in this research |
| Detector evaluation | Frozen method tested on independent plates/sessions; split/merge errors and calibration uncertainty reported | Attach actual integrated evidence; pending in this research |
| Collector validation | Held-out I–V/power residuals assessed against preregistered measurement and model uncertainty | Attach actual integrated evidence; pending in this research |
| Interface/accessibility | Critical workflows work by keyboard; readable focus/errors/tables; responsive panels and labels verified | Attach actual integrated evidence; pending in this research |
| Performance | Time/memory and cancellation measured on a stated reference device using realistic datasets | Attach actual integrated evidence; pending in this research |


Every gate should name the tested runtime/browser/device/input fixture and typed tolerance. A global1e−6 is inappropriate for differently scaled/unit-bearing outputs. More precision in a discrete residual is not automatically more scientific validity. Synthetic test reports must remain labelled synthetic even if their metrics are perfect.

## Institutional deliverables, funding and realistic programme

Prepare one primary endpoint, one falsifiable hypothesis, a repaired release and honest change log, a five-minute pinned replay, scoped numerical evidence, a data/rights/calibration plan, responsibilities and effort, preregistered split, and background-IP/publication topics. HBF contributes implementation and reproducibility; the host must actually agree to supervision, access and independent evaluation. A university name or a public research page is not an agreement. No emails, data requests or applications were sent.

The Germany/EU routes in the plan remain conditional on legal form, location, stage, affiliation and host. The official pages were revisited 11 September 2026: TransferBONUS lists entry up to €7,500 at 100%, standard up to €15,000 at 70%, and digitalization up to €45,000 at 70%. Those are programme caps, not an HBF award or a whole-programme budget. The actual eligible project must be reviewed before start. [TransferBONUS](https://www.ibb-business-team.de/transfer-bonus/).

EXIST Gründungsstipendium is submitted through the host institution; programme stage and company timing must be checked with that host. Forschungstransfer is a later research-transfer candidate requiring the appropriate research foundation, team and proof-of-principle context; an ideal web model is not measured converter proof. [EXIST Gründungsstipendium](https://exist.de/programm/gruendungsstipendium/antragsstellung/), [EXIST Forschungstransfer](https://exist.de/programm/forschungstransfer/antragsstellung/).

Pro FIT's AI/ML guidance specifically requests the objective, success criteria and extent/origin of training data; funding type varies by innovation phase. This supports collecting a defensible dataset before advertising AI as an achievement. [IBB Pro FIT](https://www.ibb.de/de/foerderprogramme/pro-fit-projektfinanzierung.html).

EIC Pathfinder Open 2026 closed 12 May 2026; Challenges is listed for 28 October 2026. Open normally requires at least three independent entities with the stated country pattern. No direct fit to Challenges 2026 is established, and no application or eligibility probability can be inferred. [EIC Pathfinder](https://eic.ec.europa.eu/eic-funding-opportunities/eic-pathfinder_en), [Challenges 2026](https://eic.ec.europa.eu/eic-funding-opportunities/eic-pathfinder/eic-pathfinder-challenges-2026_en).

### Twenty-four months and decision gates

These are programme milestones, not actions completed by adding software. Preserve the existing four broad collaboration phases and use the following as a concrete appendix. Experiments depend on a funded, supervised facility pathway.


| Period | Work | Reviewable deliverable | Proceed only if |
| --- | --- | --- | --- |
| Month 1 | Repair release/domain/status defects; select one flagship primary endpoint | Corrected release, regression evidence, concise problem statement | Another person can run the check and understand one result without interpretation errors |
| Months 2–3 | Form host/data relationship; establish contracts and image/electrical ingestion | Partner work-package draft, data/rights plan, typed schemas, existing-model baselines | Actual data or a credible scheduled acquisition path exists |
| Months 4–6 | Calibrated pilot analysis; connected power/particle records; formal evaluation design | Pilot dataset, frozen baseline, reproducible package, funding-ready work plan | Calibration is adequate and the proposed endpoint is identifiable/measurable |
| Months 7–9 | Build the first measurement workflow with the facility; implement the chosen robust analysis | Initial non-fusion measurement records and comparison plots | Readings have traceable uncertainty and the model boundary is clear |
| Months 10–12 | Complete first held-out validation and reference-model comparison | First validated subsystem result or documented model failure | Improvement is demonstrated against a baseline; unresolved effects are listed |
| Months 13–15 | Refine the strongest method; add correlated uncertainty and a second campaign | Prediction-interval evaluation and independent session results | Method remains useful when acquisition conditions change |
| Months 16–18 | External rerun or second-site test; limited experiment-selection study | Independent evaluation report and preregistered selection comparison | A partner can reproduce the key result without the original developer operating the workflow |
| Months 19–21 | Consolidate source/transport/detector/electrical links where validated | Integrated research package, methods manuscript/data release where permitted | Every headline is traceable to its supported domain and evidence |
| Months 22–24 | Final independent review, adoption testing, follow-on proposal | Public/controlled release, documented limits, next-stage decision | Follow-on scope is justified by evidence rather than visual ambition |


Stop/narrow rules remain binding: no usable data by Month 3 means a source/solver/data-infrastructure pilot; no independent detector-response calibration means no species/energy inversion; adequate baseline performance means study reliability or calibration transfer before adding ML complexity; unexplained collector residuals mean localize discrepancy before optimizing; no external reviewer/adoption by Month 12 means narrow to one reproducible package; a changed scaling regime means a separate physics question.

The plan's budget is illustrative arithmetic: 24 engineering person-months at €8,000, 18 research at €9,000, 12 instrumentation at €9,000, 6 senior review at €10,000, plus €50,000 facility/calibration,€40,000 equipment and€10,000 compute/software/archive gives €622,000; 15% contingency gives €715,300. It is not salary research, a quotation, eligibility assessment, grant expectation or reactor construction cost. Actual in-kind host contributions cannot also be counted as cash expenditure. The initial 6–12 person-month pilot arithmetic (€36,000–€120,000 before facilities at the stated illustrative rates) needs real costs and access before approval.

### First thirty days


| Days | Action | Done means |
| --- | --- | --- |
| 1–5 | Reproduce and repair B01–B04 in the source repository | Matching-release hosting check, correct inclusive temperature boundary, consistent run/export metadata and status |
| 6–10 | Promote heating closure; fix comparator units and stage labels | One result view shows electrical boundary, required heating and assumptions; units are explicit |
| 11–15 | Choose AlphaTrack or measured collector response as the primary pilot | One-page hypothesis, baseline, observable, necessary data and failure criterion |
| 16–20 | Produce a sample typed package and a short independent replay | A second person can reproduce the selected numerical/analysis output from documented inputs |
| 21–25 | Prepare the host/data work package | Named expertise needed, data/calibration list, contribution table, effort and permitted outputs |
| 26–30 | Review with a prospective scientific host and refine the scope | Documented feedback and a concrete next experiment or data task; no implied endorsement |


## Explicitly postponed scope

The original plan postpones an autonomous reactor controller, a generic chatbot portrayed as scientific validator, a broad private-data marketplace, a new general-purpose PIC/MHD solver, full commercial reactor cost forecasting, and high-fidelity 3D expansion without measured uses. F19 free-boundary/stability adapters are specialist/question-gated, not permission to silently build an unvalidated general solver. F18 collaboration requires actual users/permissions; a local embargo field is not enforced access control.

The candidate research contributions remain narrow hypotheses: cross-site calibration transfer, joint identifiable source/detector/collector inference, energy-origin composition that demonstrably prevents errors, measured discrepancy localization, cost-aware experiment choice, and validity-aware surrogates. Each requires a focused prior-art comparison and a falsifiable endpoint. Generic particle tracing, segmentation, Monte Carlo, active learning, provenance packages and net-power dashboards are established elsewhere. Public reviewed evidence did not establish global novelty.

## Complete plan coverage and artefact semantics

requirements-draft.json retains every source heading segment and full requirement text where applicable, plus B01–B09, F01–F20, C01–C30, all ten release gates, eighteen data catalogue entries, three flagships, all nine 24-month review periods, the first 30 days, stop rules, budget and postponed scope. Each pending field is intentional until actual integration or experimental evidence is attached.

source-register.json contains exact URLs, owner classes, item rights, observed versions, retrieval statuses and limitations for 46 sources/prospective data families, plus hashes of ten supplied local evidence files. A failed repository fetch is a failed retrieval, not proof that code is unavailable or unlicensed. A paper-stated public reproduction link is not a completed data acquisition. Development documentation labelled latest was not adopted as an installed dependency.


| Plan area | Coverage in this review | Evidence/achievement boundary |
| --- | --- | --- |
| 1 Decision | Correctness -> connected ledgers -> calibrated detector/collector -> uncertainty | No whole-reactor endorsement |
| 2 Audit | Historical observations and coverage limits retained | No fresh all-browser/GPU/security certification |
| 3 B01–B09 | Nine specific repair requirements and acceptance | Integration evidence pending |
| 4 Physics | Eight interpretation domains and critical equations | V1/V2 distinguished from measurement |
| 5 Existing labs | Preserve existing advanced capabilities; twenty additions extend them | No renamed duplication as new contribution |
| 6 Established software | Fifteen software/data-pattern families reviewed across source register | Docs are not executed benchmark results |
| 7 Features/flagships | All 20 ranked additions, three experiments, six novelty hypotheses, postponed scope | Training/measurement/novelty remain gated |
| 8 Data | All 18 sources and eight-step ingestion path | Public access distinct from permitted reuse |
| 9 Architecture | Thirty traceable contracts and package semantics | No fake backend/control claims |
| 10 Validation | All 14 historical fixtures, 10 gates, V0–V5 | No fabricated new pass or physical validity |
| 11 Partners | Review package, contribution split and bounded host request | No partner agreement/contact made |
| 12 Funding | Six current route/theme pages and eligibility caveats | No award/application/odds claims |
| 13 Programme | Nine milestone periods, stop rules, budget and 30-day sequence | Future programme pending facility/funding |
| 14 Evidence | Source segments/hash register and remaining questions | Exact retrieval and provenance scope retained |


## Primary-source index

Every link below has a distinct retrieval status. Detailed rights and intended use are recorded in the JSON.


| ID | Source | Retrieval status |
| --- | --- | --- |
| D01 | [Supplied HBF Sikora–Weller derived table](https://hbfnetwork.org/data/pb11-sikora-weller-2016.csv) | supplied-local-table-read |
| R01 | [AME2020 and NUBASE2020 official files](https://www-nds.iaea.org/amdc/) | documentation-retrieved |
| R02 | [LiveChart CSV API guide](https://www-nds.iaea.org/relnsd/vcharthtml/api_v0_guide.html) | documentation-retrieved |
| R03 | [IAEA EXFOR](https://www-nds.iaea.org/exfor/) | landing-retrieved-no-record-content |
| R04 | [IAEA IBANDL](https://www-nds.iaea.org/ibandl/) | redirected-landing-no-record-content |
| R05 | [NIST stopping-power and range tables](https://www.nist.gov/pml/stopping-power-range-tables-electrons-protons-and-helium-ions) | documentation-retrieved |
| R06 | [NIST copyright and licensing policy](https://www.nist.gov/open/copyright-fair-use-and-licensing-statements-srd-data-software-and-technical-series-publications) | documentation-retrieved |
| R07 | [CIAAW isotopic abundance table](https://www.ciaaw.org/isotopic-abundances.htm) | documentation-retrieved |
| R08 | [IAEA FENDL](https://www-nds.iaea.org/fendl/) | documentation-retrieved |
| R09 | [OPEN-ADAS terms](https://open.adas.ac.uk/terms-and-conditions) | documentation-retrieved |
| R10 | [CODATA full constants listing](https://physics.nist.gov/cuu/Constants/Table/allascii.txt) | text-data-retrieved-not-saved-as-raw-file |
| P01 | [CR-39 D–D deep-learning study](https://arxiv.org/abs/2503.06932) | documentation-retrieved |
| P02 | [Xie bremsstrahlung fitting](https://arxiv.org/html/2404.11540v2) | documentation-retrieved |
| P03 | [Adiabatic direct energy conversion](https://arxiv.org/abs/2410.15485) | documentation-retrieved |
| P04 | [p–B reactor breakeven](https://arxiv.org/abs/2310.18508) | documentation-retrieved |
| P05 | [p–B alpha ash](https://arxiv.org/abs/2502.13300) | documentation-retrieved |
| P06 | [IAEA Fusion Data Lake report](https://arxiv.org/abs/2604.01797) | documentation-retrieved |
| P07 | [FreeGSNKE/Fiesta MAST-U equilibrium comparison](https://arxiv.org/html/2407.12432v5) | documentation-retrieved |
| P08 | [NRL Plasma Formulary2023](https://www.nrl.navy.mil/Portals/38/PDF%20Files/NRL_Plasma_Formulary_2023.pdf) | documentation-retrieved |
| S01 | [ASCOT5 documentation](https://ascot4fusion.github.io/ascot5/) | documentation-retrieved |
| S02 | [WarpX parameter documentation](https://warpx.readthedocs.io/en/latest/usage/parameters.html) | documentation-retrieved |
| S03 | [COMSOL Particle Tracing](https://www.comsol.com/particle-tracing-module) | documentation-retrieved |
| S04 | [SIMION space charge](https://simion.com/info/space_charge.html) | documentation-retrieved |
| S05 | [PROCESS](https://ukaea.github.io/PROCESS/) | documentation-retrieved |
| S06 | [bluemira](https://bluemira.readthedocs.io/en/latest/) | documentation-retrieved |
| S07 | [OpenMC](https://docs.openmc.org/en/stable/) | documentation-retrieved |
| S08 | [Geant4](https://geant4.web.cern.ch/docs/) | documentation-retrieved |
| S09 | [ilastik documentation](https://www.ilastik.org/documentation/) | documentation-retrieved |
| S10 | [Cellpose documentation](https://cellpose.readthedocs.io/en/latest/) | documentation-retrieved |
| S11 | [SALib](https://salib.readthedocs.io/en/latest/) | documentation-retrieved |
| S12 | [openPMD API](https://openpmd-api.readthedocs.io/en/latest/) | documentation-retrieved |
| S13 | [IMAS-Python](https://imas-python.readthedocs.io/en/latest/) | documentation-retrieved |
| S14 | [RO-Crate1.2 specification](https://www.researchobject.org/ro-crate/specification/1.2/) | documentation-retrieved |
| F01 | [TransferBONUS Berlin](https://www.ibb-business-team.de/transfer-bonus/) | documentation-retrieved |
| F02 | [EXIST Gründungsstipendium application](https://exist.de/programm/gruendungsstipendium/antragsstellung/) | documentation-retrieved |
| F03 | [EXIST Forschungstransfer application](https://exist.de/programm/forschungstransfer/antragsstellung/) | documentation-retrieved |
| F04 | [IBB Pro FIT](https://www.ibb.de/de/foerderprogramme/pro-fit-projektfinanzierung.html) | documentation-retrieved |
| F05 | [EIC Pathfinder](https://eic.ec.europa.eu/eic-funding-opportunities/eic-pathfinder_en) | documentation-retrieved |
| F06 | [EIC Pathfinder Challenges 2026](https://eic.ec.europa.eu/eic-funding-opportunities/eic-pathfinder/eic-pathfinder-challenges-2026_en) | documentation-retrieved |
| X01 | [WarpX licence URL attempted](https://github.com/BLAST-WarpX/warpx/blob/development/LICENSE.txt) | retrieval-failed |
| X02 | [ASCOT5 licence URL attempted](https://github.com/ascot4fusion/ascot5/blob/main/LICENSE) | retrieval-failed |
| X03 | [FreeGSNKE repository URL attempted](https://github.com/PlasmaFAIR/FreeGSNKE) | retrieval-failed |
| X04 | [Xie linked code repository](https://github.com/hsxie/brem) | linked-from-paper-not-retrieved |
| D16 | Instrument manufacturer calibration certificates | not-acquired |
| D17 | University raw plates and collector campaigns | not-acquired |
| D18 | Supplier/material characterization and partner measurements | not-acquired |