SCPN Control

SCPN Control is a neuro-symbolic fusion-control package for turning controller ideas into auditable, replayable, evidence-bound software artefacts. It combines stochastic Petri nets, spiking neural control, robust and nonlinear MPC, phase-dynamics monitoring, differentiable physics facades, public-data validation gates, and runtime safety contracts.

Plain-language summary

A fusion control team needs more than a simulator. It needs a way to answer five questions before any controller moves toward hardware:

1. What physical contract is this controller allowed to claim?
2. Which data, reference code, or hardware evidence supports that claim?
3. Does the control logic satisfy bounded safety invariants?
4. Can the runtime fail closed when evidence, authentication, or configuration is missing?
5. Can another lab replay the result and see the same admission decision?

SCPN Control is built around those questions. It is not a replacement for a facility PCS, EFIT deployment, TRANSP workflow, or external gyrokinetic code. It is the controller-facing evidence layer that makes those integrations safer to discuss, test, fund, and review.

Current release focus

The current release focus is native execution evidence and formal-runtime claim discipline. SCPN Control now separates Python-orchestrated campaigns from fused Rust/PyO3 campaigns, reports formal-verification coverage explicitly, and uses digest-bound AOT certificate evidence for the hot path. Non-isolated workstation timings remain local-regression evidence; production benchmark claims still require isolated cores, host-load metadata, governor context, and target-hardware or PREEMPT_RT admission.

Main surfaces

Surface	Purpose
Petri-net to SNN compiler	Build neuro-symbolic controllers from explicit places, transitions, and firing contracts.
Formal verification	Produce bounded reachability, marking, liveness, CTL/LTL, and certificate-bundle evidence for control logic.
NMPC and robust control	Prototype constrained plasma-control policies with strict solver lifecycle and admission checks.
Differentiable physics facades	Expose JAX-backed controller-tuning paths without claiming facility fidelity before admission.
Digital twin and replay	Couple controller logic to replayable plant and external-simulator metadata.
Public-data validation	Convert and evaluate public datasets while keeping predictive claims blocked until full evidence exists.
Runtime security	Harden WebSocket, native-code, config, and artefact-loading boundaries for fail-closed operation.
Documentation and release evidence	Publish enough context for users, collaborators, reviewers, and funders to understand the product boundary.

What the software is for

SCPN Control is for the stage where a fusion-control idea needs to become a reviewable software artefact. The package helps you answer: what controller was run, what physics/replay contract fed it, what assumptions were declared, what validation artefact was produced, what checksum binds that artefact, and what stronger claims are still blocked.

The core product is not a single algorithm. It is the control-evidence workflow: controller construction, bounded verification, physics/replay integration, benchmarking, report generation, strict admission, documentation, and release tracking. That workflow is useful to research groups, startups, facility teams, and funders because it turns ambiguous demonstrations into auditable evidence.

Application and value map

Application	Immediate value	Evidence still required for stronger claims
Controller concept review	Prototype SPN/SNN, NMPC, robust, phase, and digital-twin controllers with explicit assumptions	HIL replay, target-hardware timing, independent safety review
Neural equilibrium preparation	Prepare public MAST EFM datasets, launch reports, and result templates without hiding blockers	Executed full-output training, holdout metrics, latency/GPU-cost reports, EFIT/P-EFIT admission
Differentiable control tuning	Use JAX transport/equilibrium paths and gradient evidence for optimisation experiments	External physics validation and target-hardware timing evidence
Gyrokinetic validation	Persist CPU/GPU backend parity and local bounded reports	Real GENE, TGLF, GS2, CGYRO, or QuaLiKiz comparison artefacts
Safety-case infrastructure	Emit formal proof and certificate bundles for bounded controller logic	Facility hazard analysis, independent V&V, requirements traceability
Collaboration and financing	Show exactly what data, compute, and hardware evidence are missing	Funded acquisition, external-code runs, facility-data permissions

Visual overview

Fast start

pip install scpn-control
scpn-control demo --steps 1000
scpn-control validate

For a guided path, start with Onboarding, then run the First Steps tutorial, then choose a workflow from Use Cases.

Evidence boundary

SCPN Control publishes many validators and reports. A passing bounded validator means the repository contract for that surface passed. It does not mean that a facility, regulator, ITER safety case, or external physics authority has accepted the result.

A full-fidelity public claim requires all required artefacts for that surface: measured-shot or documented public-reference data, external-code or independent reference where applicable, exact unit contracts, checksums, declared tolerances, target hardware timing where relevant, and human review.

Support and collaboration

The project needs collaborators and financing for target-hardware benchmarks, external-code validation, public-data curation, cloud GPU campaigns, safety-case review, and facility replay access. See Compute Validation Funding for concrete budget categories and why each category matters.