Skip to content
Technology How it works Breeder — Hyperion Burner — Aegis Burner — MetroVolt AI-Native Architecture Magnets Fuel cycle Safety Roadmap
Solutions AI & Data Centers Defense & Government Grid & Baseload Neutron Detection Quantum
Learn Technical Library
Proof Publications Whitepapers Technical Library Open Science & Reproducibility The Honest Gates
Company About / Mission Leadership Environment Health & Safety Investors Careers Press Contact
3D Model
AI Architecture › L5 · Applications & Copilots
L5 · Applications & Copilots

Copilot Audit and Provenance

Every copilot request, reasoning trace, citation, proposal, and authorization decision is retained immutably so any output can be reconstructed and reviewed.

THE STACK · click to jumpL7Ecosystem & StrategyL6Experience & VisualizationL5Applications & CopilotsL4OrchestrationL3Twin Modeling & AIL2Data FabricL1Control PlaneL0Foundation▲tlmctl▼L5 · APPLICATIONS & COPILOTSAgentic copilots that reason over the machine.1Plasma Copilotscenario design2Engineering Copilotsubsystem analysis3Operations Copilotrunbooks & procedures4Agentic Toolsbounded action-taking5Knowledge BaseRAG over the fabric6Guardrailssafety-boundedMACHINE TIEReads the twin and fabric; proposes actions that route through L4.KRONOS FUSION ENERGYAI-NATIVE S.M.A.R.T. GENERATORAPPLICATIONS & COPILOTSSHEET 07REV. 2026-08L5 · AI-NATIVE STACK
L5 · Applications & Copilots — its place in the stack (left, click any layer) and its internal components (right). Telemetry rises; control descends.

Everything is on the record

A copilot that advises on a nuclear machine must be fully auditable. Kronos retains, for every request, the complete record: the assembled context (with the twin-state snapshot and prompt version), the reasoning trace of tool calls and observations, the retrieved sources with their provenance ids, the composed answer with its citations, any proposal submitted, and the L4 authorization and human-approval decisions.

The audit record

text
audit_record:
  request_id, timestamp, operator, copilot, machine
  context: {prompt_version, twin_state_snapshot_id, budget}
  trace:   [(thought, tool_call, observation), ...]
  sources: [provenance_id, ...]
  answer:  {text, citations[], uncertainty}
  proposal: {action, L4_decision, human_decision} | null
  eval:    {grounding_score, calibration}   # if sampled
  -> append-only, hashed, linked to L2 lineage

Why this depth

The record is append-only and hashed, linked into the L2 data-lineage and provenance systems, so it cannot be quietly altered. Because prompt version, model version, twin version, and retrieval snapshot are all captured, any answer is reproducible: re-running the recorded context should yield the recorded answer, which is itself a check on determinism.

Confidentiality preserved in the record

Audit depth does not override confidentiality. The record respects the same tier controls as retrieval: internal-confidential sources are referenced by provenance id and access-scoped, never copied into a lower-tier context. This keeps the full chain reviewable by authorized people while honoring the firewall between confidential and public material.

The audit record is the ground truth for accountability, evaluation, and improvement. It closes the loop with continuous evaluation (production cases become test cases) and underpins human oversight (an operator can always answer 'why did it say that').

Content reviewed August 2026 · design-and-simulation stage