AI agent reliability · AI security
Agent Review Control
The public Contract + Evidence + Verdict architecture for controlling and reviewing AI coding-agent work.
StatusPublic architecture and dogfood evidence; active deterministic implementation remains private.
ContractFrozen before work
EvidenceBound to exact change
VerificationIndependent and fail-closed
DecisionHuman-controlled
01 / Problem
Why this system exists
Coding agents can execute successfully while drifting from the approved assignment or presenting evidence that does not belong to the exact change under review.
02 / Architecture
How state moves
- 01Frozen contract
- 02Execution controls
- 03Bound evidence
- 04Independent verification
- 05Trust Brief
- 06Human approval
03 / Decisions
Engineering choices
- Freeze approved intent and boundaries before implementation begins.
- Treat agent claims as untrusted until supported by attributable evidence.
- Keep verification independent from the agent that produced the work.
- Track evidence through pending, current, stale, and refreshed lifecycle states.
04 / Boundaries
What this does not claim
- ARC does not prove code correctness, safety, or semantic truth.
- Private doctrine, customer evidence, operational controls, and source remain private.
- CI, security tools, human review, and ARC evaluate different properties.