Compute and validation collaboration¶
SCPN Control is research infrastructure for bounded fusion-control experiments. It includes formal Petri-net checks, differentiable physics surfaces, neural transport and equilibrium facades, digital-twin contracts, and strict evidence admission gates. External compute, data, hardware, and technical review can strengthen the evidence available through those public interfaces.
Claim boundary
Sponsorship, credits, data access, or collaboration do not create facility deployment evidence, safety approval, or a quantitative scientific claim. Evidence is admitted only when the corresponding public validator passes.
Evidence interfaces¶
Collaborators can supply reviewable artifacts in these forms:
| Evidence class | Reviewable contribution |
|---|---|
| Public datasets | Stable source URI, licence, size, checksum, retrieval date, feature/target schema, and units |
| External-code comparison | Input decks, raw output, parsed output, code/version identity, units, tolerances, and digests |
| Model training | Dataset manifest, preprocessing contract, seeds, hardware metadata, weights, holdout predictions, and uncertainty |
| Hardware timing | Exact source revision, build identity, host and accelerator metadata, sample definition, scheduler/load context, and raw samples |
| Facility replay | Data-policy authority, immutable shot/signal identity, calibration and units, replay configuration, output metrics, and claim boundary |
| Independent review | Reproducer, observed result, environment details, and a clear distinction between confirmation and disagreement |
Failed comparisons are useful evidence when their inputs and outputs are preserved. They remain failures; they are not rewritten into positive claims.
Public dataset references¶
The repository contains normalized acquisition manifests for these public neural-transport sources. Large payloads are not committed to Git.
| Dataset | DOI | Files | Published payload size |
|---|---|---|---|
| QLKNN10D training set | 10.5281/zenodo.3497066 |
5 | 32,080,102,848 bytes |
| QuaLiKiz v2.6.2 JET spectra | 10.5281/zenodo.7418108 |
1 | 29,655,790,232 bytes |
| QLKNN11D training set | 10.5281/zenodo.8017522 |
46 | 247,952,755,894 bytes |
Possession of these payloads does not admit a neural-transport claim. The public claim report also requires bound preprocessing, weights, holdout predictions, metrics, provenance, and an admitted evidence class.
Current claim boundaries¶
- Neural transport uses an analytic critical-gradient fallback unless admitted weights and reference evidence are supplied.
- Neural-equilibrium assets are synthetic pretraining inputs until an identical-input EFIT/P-EFIT holdout report passes.
- External gyrokinetic agreement is absent until a strict matched-input report binds real TGLF, GENE, GS2, CGYRO, or QuaLiKiz output.
- Synthetic disruption fixtures exercise data and prediction plumbing; they do not establish a measured facility ROC.
- Local and CI timing observations do not establish deterministic PCS, HIL, or production behavior.
The generated physics traceability report is the public authority for component-level claim status. The validation guide explains evidence classes and the benchmark guide explains timing context.
Contact¶
Compute sponsorship, cloud credits, storage, facility-data collaboration, external-code artifacts, target-hardware measurements, and independent review are welcome. Contact protoscience@anulum.li with the evidence or resource class, applicable licence or access policy, and the public interface you want reviewed.
Public issue trackers #46 through #53 provide neutral contribution intake by evidence domain. They do not disclose or prescribe an internal work sequence.