Integrity · Capture
We intercept at the boundary. Not inside the model.
Internal reasoning is unobservable. What we can capture mechanically is every tool call, every byte returned, every config and prompt — plus what the agent explicitly declares.
Passive vs active capture.
Green = observed on the wire. Amber = what the agent asserts and must prove.
Observed (passive)
activity.start/end · entity.observed · task.assigned · config.observed · message.final — captured without the agent writing integrity events.
Declared (active)
artifact.published · claim.submitted — explicit claims with provenance. Zero trust until resolved against observed nodes.
Pipeline
Append-only journal → deterministic graph.
1Capture busAdapters normalize events
2Journal + CASSHA-256 content store
3MaterializerPure fold — no LLM
4VerifierPolicy per vertex type