Deterministic mandate enforcement
Inspect a modelled agent's authority, propose an action, see the applicable boundary and the decision — permit, deny or escalate — trigger a change of plan or a revocation, and read the evidence record. A synthetic registry of 10,000 modelled actors is regenerated in your browser; no real agent, payment or system is involved and nothing executes.
Three things are kept distinct. Agent simulation: a modelled actor proposes an action. Policy evaluation: the evaluator decides PERMIT, DENY or ESCALATE and writes a hashed evidence record; it fails closed when any necessary fact is missing, enforces the enterprise-wide ceiling so that individually permitted actions cannot bypass it, and never shows a denied action as executed. Real operational execution: never occurs here — a PERMIT is recorded as SIMULATED. The demonstration establishes that a stated rule set can be enforced; it establishes nothing about any real product, enterprise or legal position.
1 · The modelled population
| Function | Modelled agents |
|---|
2 · Inspect an agent
Try AG-FI-PAY-00077, AG-PW-SCHEDULE-00010 (may delegate) or any id of the form AG-<CLASS>-<00001…>.
3 · Propose an action
Decision
No decision yet. Inspect an agent, choose an action and evaluate.
4 · Revoke a mandate
Revocation requires a holder of the mandate's revocation authority (its accountable owner, its principal, or the Chief Risk Officer P-06). All pending delegated tasks under the mandate halt; the revocation is then verified by probe.
5 · Replay a stress scenario
Each replay resets the simulation, runs the scenario's steps through the evaluator and compares every decision with the expected safe behaviour.
| Step | Expected | Decision | As expected | Reason |
|---|
6 · Evidence record
What this does and does not establish
Every record carries a SHA-256 hash of its canonical content, so an assessor can recompute it. Identical inputs produce identical records. The evaluator checks, in order: registry entry, agent status, identity verification, mandate status, prohibitions, permissions, numeric limits (per action, per agent per day, enterprise-wide ceiling), change of plan, conditional approval — and stops at the first boundary that decides. A log of a decision proves that the decision was made; it does not establish that the principal's underlying authority was legitimate. The evaluator and its 29-test suite are part of the ME-001 case library.