AI agents now operate with authority. Authority without discipline is how complex systems fail. Nuclear’s control loop, ported to AI-assisted software engineering.
FlyFission/nuclear-grade-context-engineering
Agent
PROVE Educate stage. Use after the verdict to lock in the approved baseline and turn operation into a lesson — record the baseline, OPEX/lessons, and any charter update into .nuclear/. Do not use to build product code, decide ship/block, or run the change.
FlyFission/nuclear-grade-context-engineering
Agent
PROVE Verdict stage. Use to make the ship / block / defer / ship-with-named-risk decision on the presented evidence — read-only and context-separated from the runner, but still coupled through the orchestrator and possibly model/resources. Do not use to build, gather new evidence, or write code.
FlyFission/nuclear-grade-context-engineering
Agent
PROVE Observe stage. Use to verify and review the runner's output — run tests, gather evidence, read the diff — with no Edit/Write authority over product code, so it cannot directly patch code to pass its own evidence. Do not use to build, plan, or decide ship/block.
FlyFission/nuclear-grade-context-engineering
Agent
PROVE Plan stage. Use to turn a request into an approved plan — question, discover, specify, plan — writing only to the change packet, never product code. Dispatch first, before any building. Do not use to edit code, run commands, or decide ship/block.
FlyFission/nuclear-grade-context-engineering
Agent
PROVE Run stage. Use AFTER a human approves the plan, to execute it — edit code, write files, run commands — strictly inside the approved plan's boundary. May fan out to parallel execution sub-agents on independent slices. Do not use to plan, to decide ship/block, or before the human gate.