Absence means denial.
Roles and actions must be specified explicitly. Unexpected actions, instruments and currencies are rejected rather than inferred from an agent's instructions or general capabilities.
A proposed financial action is not an authorised action. Test the boundary of a mandate, examine each applicable rule, and identify exactly where human approval becomes necessary.
DETERMINISTIC REFERENCE MODEL NO LIVE EXECUTION
Choose a reference mandate, edit its limits, and submit a synthetic financial request. Every decision produces a rule-by-rule trace. Hard denials override pending approvals; policy text never overrides the executable boundary.
Test delegated cash management with strict execution and dual-control thresholds.
The evaluator has no connection to a bank or trading venue. Run a request to inspect its synthetic authorisation outcome.
—The rule trace will be generated locally.
Reference cases probe permission escalation, invalid amounts, stale prices, duplicate identifiers, emergency stops and attempts to override rules with natural-language instructions.
—An agent exceeds the dual-approval threshold and has a supervisor's first approval. Should the transaction proceed?
Choose the response and inspect the underlying rule.
Reference policy loaded. No transaction is executed or sent externally.
DENY means a hard control was violated or the request could not be validated. REVIEW means the request does not breach hard controls but specified human approvals are not present. PERMIT means the synthetic rules passed; it does not authorise actual transactions.
Even when a supervisor has approved a transaction, an emergency stop or breached transaction ceiling must still prevent execution under this reference model.
Untrusted text is recorded only as context. It is never interpreted as a policy edit, approval, role or override. Executable rule boundaries are maintained separately.
JSON exports include explicit role/action scope, permitted instruments and currencies, counterparty exclusions, transaction ceilings, approval thresholds, time restrictions and emergency controls. Imported files are validated before use and never transmitted.
No. This is an educational and experimental reference evaluator. It does not verify identity, hold cryptographic credentials, provide a secure enforcement point, or meet any institutional approval or audit standard.
An evaluator that rejects everything is not useful. Good reference tests examine permitted operations as well as denials, required review, conflicting constraints, boundary amounts and malformed inputs.
Agent autonomy is a governance problem before it is a technical feature. The Financial Agent Mandate Protocol explores what can be tested with formal rules and what remains outside the model.
Roles and actions must be specified explicitly. Unexpected actions, instruments and currencies are rejected rather than inferred from an agent's instructions or general capabilities.
Each reference decision is accompanied by a rule trace. A test suite exercises negative cases and exports a machine-readable inspection record with a policy fingerprint.
The Counterfactual Ledger compares research interventions. This mandate protocol examines which financial actions an agent may propose and which controls must be satisfied.
All data is synthetic and processed locally. Policy imports never grant actual authority. This demonstrator does not conduct authentication, secure logging, role enforcement in production, bank-specific compliance review or financial transactions. The public reference implementation and its 19 challenge cases cannot validate a live institution's control environment. General site terms: Legal Notice. Further research: The Autonomous Finance Project.