Verifiable AI · release control
Agent Integrity
A deterministic release-control engine that checks AI responses against recorded evidence, decisions, and signed receipts before release.
StatusPublic alpha v0.1.0-alpha.2; source installation is available while npm publication remains pending.
OutcomesPASS · REVIEW · BLOCKED
ReleaseExact checked bytes only
ReceiptsEd25519 authenticated
StatusPublic alpha
01 / Problem
Why this system exists
Agent applications often let the same model gather evidence, interpret policy, write an answer, and declare its own work valid. That makes unsupported claims, hidden contradictions, stale receipts, and post-verification response changes difficult to catch reliably.
02 / Architecture
How state moves
- 01Trusted policy
- 02Source collection
- 03Claim + evidence envelope
- 04Deterministic verifier
- 05PASS / REVIEW / BLOCKED
- 06Exact response release
03 / Decisions
Engineering choices
- Keep verification deterministic and independent of model-provider judgment.
- Bind evidence, decisions, and response sections to exact bytes and digests.
- Fail closed when input is malformed, verification fails, or released bytes change.
- Authenticate producer receipts and enforce atomic single-use consumption on one protected host filesystem.
04 / Boundaries
What this does not claim
- A PASS proves configured consistency and tamper-resistance checks passed; it does not prove objective truth, safety, or correctness.
- The verifier checks declared dependencies and cannot infer evidence or decisions the application omitted.
- The current alpha is tested for server-side Node.js and is not published to npm yet.