MUJ428 Trust Layer
Human overview · Verify before act

MUJ428 Trust Layer

MUJ428 Trust is independent trust infrastructure for consequential AI-agent actions. It is designed to verify that the action an agent was authorized to take is the action that actually occurs.

Open the live MUJ428 Trust site

What MUJ428 Trust does

AI agents can purchase, pay, approve, write, delegate, release, accept, and commit on behalf of people and organizations. MUJ428 adds an integrity layer between authorization and execution so approved intent can be checked against actual effect.

The model is simple: verify before act, then verify what happened. MUJ428 can bind exact intent, action arguments, authority, evidence, freshness conditions, and execution premises into a Trust Receipt. A Verified Effect records the resulting state.

MUJ428 does not take custody of the underlying transaction principal. The buyer, seller, agent, platform, or payment system remains responsible for the actual transaction or side effect.

The problems it targets

Intent drift

An agent can begin with an approved goal but later invoke a materially different action or arguments. MUJ428 makes the approved action explicit and checkable.

Stale evidence

A decision can become unsafe when evidence, policy, resource state, or preconditions change. MUJ428 treats freshness as part of the trust decision.

Replay and duplicate effects

Consequential actions need strong handling of retries, idempotency, and uncertain outcomes. MUJ428 includes replay-aware effect-state concepts.

Weak proof after execution

Knowing what an agent intended is not enough. MUJ428 also emphasizes a machine-readable record of what effect actually occurred.

Built for humans and machines

The public MUJ428 Trust site includes a human product interface plus REST, MCP, A2A, OpenAPI, Agent Card, and other machine-readable discovery resources. For people, the core point is straightforward: MUJ428 is designed to make consequential AI-agent actions more inspectable, deterministic, and verifiable.