Intake and trust classification.
Adapters snapshot sources, record hashes and provenance, classify trust, normalize records, and place changes into queues. Change intake and command execution remain distinct concerns.
Trace-first acquisition and verification.
Candidate claims and relationships are proposed from code, contracts, tests, traces, diffs, artifacts, and cited worker support sets. Acquisition records provenance and support edges before graph structure is trusted. Agent-generated structure remains unverified until claim-specific evidence supports it.
Write ownership stays explicit.
The authority engine admits scoped changes to the Ratified Intent Graph. Verification processes update claim state in the Evidence-Backed Reality Graph. Reconciliation writes current orders to the Active Execution Graph. Every consequential decision and side effect appends to the Immutable Evidence and Command Ledger. No worker or evidence record writes ratified intent directly.
Execution routes locally or remotely.
A normal command enters the proxy and policy router. Lightweight permitted work can remain local; heavy builds and tests can run in isolated remote capacity. Both routes return an immutable evidence envelope before a current-epoch manager gate can perform an official side effect.
The control loop has stops.
Contradictions, missing support, stale epochs, policy denial, poisoned evidence, exhausted repair budgets, repeated failures, and non-converging plans stop or quarantine advancement. Idempotency keys prevent a replayed event from duplicating an official effect.
These are intended end-state boundaries; dated Build Records show which behavior has been demonstrated.