Seedlabs

Pre-Registered Certification Dossier Generator for ML Real-Time Controllers

A validation harness for ML-based real-time controllers (e.g., plasma control) where teams pre-register acceptance criteria before testing, then auto-generate a regulator-ready evidence pack: seeded runs with tolerance bands, per-artifact integrity manifests, and replayable verdicts.

Cross-Domain Digital Twin IntegrationCertification and licensing of AI-controlled fusion and energy systems

Concept

ML surrogates are moving into real-time control of safety-critical machines, but there is no standard artifact a licensing authority can review. The tool implements the certification pattern demonstrated in [0] as a product: a control team first writes a pre-registered, timestamped acceptance-criteria register (what performance the controller must hit, on which scenarios, with what tolerances); the harness then executes the validation matrix across pinned environments and multiple seeds, classifying every metric into tiers — deterministic quantities that must reproduce byte-identically (Tier-1) versus trained-model metrics reported as five-seed tolerance bands (Tier-2) — and emits per-artifact checksums and a replayable dossier. A regulator or third-party auditor can re-derive the verdict from the manifest rather than trusting the vendor's summary, and the same dossier doubles as the publication-grade validation record.

Why now

Paper [0] shows this entire evidence chain executed end-to-end for the KRONOS-CTRL digital-twin control architecture across two compact fusion machine classes — pre-registered acceptance criteria, fixed-seed byte-identical Tier-1 reproduction, Tier-2 tolerance bands, MD5 manifests, and adversarial audits — proving the pattern is practical today, not aspirational. The demand is documented beyond fusion: clinical AI deployment research [3] shows safety-critical domains already demand extensive upfront validation before AI tools go operational, and execution-time authorization work [5] establishes replayable verdicts and immutable decision artifacts as the emerging governance primitive that reviewers expect from AI-controlled systems.

Written by GLM · 2 Oct 2026

AI assessment

Backed by 3 papers89

A high-conviction, specialized tool that transforms a proven academic certification pattern into a commercial compliance product for high-stakes energy infrastructure.

Evidence strength
5/5
The idea is directly derived from a comprehensive reproducibility package [1] that already implements the exact workflow (pre-registration, Tier-1/2 metrics, MD5 manifests) for fusion control.
Market pull
4/5
The target customers (NRC, fusion startups) operate in highly regulated environments where the cost of a certification failure is catastrophic, creating strong urgency for standardized evidence packs.
Novelty & moat
4/5
While it shares 'verdict' concepts with the Tamper-Evident Verdict Channel, it applies them to a distinct domain (real-time physical controllers) with a unique pre-registration mechanism not present in the catalog.
Feasibility
4/5
The core logic is already demonstrated in the KRONOS-CTRL harness; the MVP requires wrapping this existing logic into a user-facing product rather than inventing new physics or AI.
Wedge clarity
5/5
The wedge is extremely sharp: a 'Certification Dossier' for ML-based plasma controllers, targeting a specific, high-value bottleneck in fusion energy licensing.
Simplicity / focus
5/5
The product is a single, focused utility—a validation harness and report generator—avoiding the 'platform' trap.

Scored by Gemma 4 · 2 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • General Atomicscompany

    Operates the DIII-D tokamak and builds plasma control systems; a standard pre-registered validation dossier would let them qualify ML surrogates for operational control and hand reviewers a replayable evidence pack.

  • U.S. Nuclear Regulatory Commissionorganization

    Now licenses US fusion plants and will receive AI-based control safety cases; a standardized, replayable dossier format makes ML controller reviews tractable for its technical staff.

  • Commonwealth Fusion Systemscompany

    SPARC's control systems will face licensing scrutiny; pre-registered acceptance criteria and tiered tolerance-band evidence streamline regulatory engagement for ML-based control components.

  • UK Atomic Energy Authorityorganization

    Must qualify increasingly AI-based control for the STEP prototype power plant; a pre-registered certification harness gives a defensible validation record for UK regulators.

  • Electric Power Research Instituteorganization

    Grid operators are adopting ML-based real-time controls with the same validation burden; a proven dossier format transfers directly to power-system reliability reviews.

Research it builds on

  1. [SUPERSEDED] Physics-Informed AI and Quantum Technologies for Certifiable Real-Time Control of Compact Fusion Generators — reproducibility deposit (code, data, figures)
    Priyanca Ford · 2026 · 360 citations
    All ideas from this paper →
  2. Execution-Time Authorization for AI Agents: A Formal Framework for Deterministic Governance Boundaries
    Edward Meyman · 2026 · 18 citations
    All ideas from this paper →
  3. Extensive upfront validation and testing are needed prior to the clinical implementation of AI‐based auto‐segmentation tools
    Justin Roper, Mu‐Han Lin, Yi Rong · 2022 · 14 citations
    All ideas from this paper →
Leave feedback
feasibility