Evidence & Reviewer Route
Do not trust the pitch. Inspect the proof boundary.
Agent Trust Gate™ is published as a reviewer-first local demonstrator. A serious reviewer can inspect the public repository, run deterministic scenarios, observe successful GatePass decisions and refusals, and examine the claims boundaries before discussing a pilot.
Exact-Action Control Model
Read the public reference model for Agent Standing, mandate, evidence, Verified Human Authority, GatePass, refusal and execution receipts.
Exact-Action GatePass
Review scoped, time-bound, action-specific proof, changed-action refusal, expiry and replay controls.
Verified Human Authority
Inspect fictional scenarios covering active identity, authority limits, separation of duties, second approval and exact-action binding.
Agent Standing
Inspect software-agent identity, accountable principal, delegation, revocation, limits and the GatePass precondition.
Reviewer Kit
Run the compact local reviewer path from a fresh clone. Expected refusals only count as success when the system actually refuses.
What a reviewer should be able to establish
- what exact action was proposed;
- which agent or workflow requested it;
- which accountable principal and current mandate were presented;
- which scope, evidence and human approval requirements applied;
- whether authority remained valid for the exact action;
- why the action was allowed, refused or left unverifiable;
- whether later simulated execution matched the allowed action.
Evidence is useful only if refusal is visible
A trust system that demonstrates only successful approvals does not show where its boundary actually is. The ATG reviewer route therefore includes changed-action refusal, expired or invalid authority, separation-of-duties failures, missing second approval and other synthetic rejection scenarios. A reviewer should be able to distinguish an allowed action from a refusal and see which control caused the difference.
This matters because the public claim is deliberately narrow: ATG is not presented as a general AI-governance platform or a guarantee that an agent is safe. It demonstrates how an exact consequential action can be checked against current authority and evidence before that action is permitted to cross an execution boundary.
Strict current boundary
The public implementation is non-production, local and synthetic. It has no live payment or settlement capability, no customer data, no production credentials, no hosted GatePass API, no live agent interception, no production key custody and no claim of regulatory approval, compliance certification or production adoption.
Commercial route
A bounded paid evaluation or design-partner discussion can begin only after the public material has been reviewed and the proposed scope is understood. Real integration would require separate design, security, legal and operational work with the organisation involved.
Read the public control model → Open the public repository → Contact the project →
Agent Trust Gate™ — Exact-Action Authority for AI Agents. Evidence before claims.