How MUJ428 Trust works
MUJ428 is designed to place a verifiable trust contract between approved AI-agent intent and a consequential real-world effect.
The execution flow
1. Define the exact action
The system identifies what is about to happen: a payment, purchase, approval, write, delegation, release, commit, acceptance, or another consequential action.
2. Check evidence and premises
Relevant evidence, policy epoch, resource versions, preconditions, authority scope, ceilings, expiry, and other current facts can be evaluated before execution.
3. Issue a Trust Receipt
The approved action and the premises supporting it are bound into a deterministic authorization artifact that can be validated by the executor.
4. Enforce at the mutation boundary
The receipt is checked at the last point where the side effect can still be prevented. A mismatch, stale premise, replay, or invalid receipt should fail closed rather than silently proceed.
5. Record the effect
After execution, a Verified Effect records the resulting state so systems can compare what actually happened with what was approved.
6. Preserve auditability
The intent, evidence, trust decision, execution conditions, and effect state form a clearer record for later review than an opaque model decision alone.
Why this is different from a simple risk score
A score can say an action looks safe. MUJ428 is aimed at a stricter question: was this exact consequential action authorized under current premises, was that authorization enforced where the effect could still be stopped, and did the resulting effect match?
That makes MUJ428 useful as a control layer around actions rather than only as a recommendation engine.
Protocol surfaces
The live service exposes human-readable pages and machine interfaces so different systems can discover the same trust model. Public surfaces include REST, MCP, A2A, OpenAPI, an Agent Card, and other discovery files.