Policy is the permission layer of a Leash identity. It tells owners, peers, and runtimes what the agent may do before the agent spends money or calls an external service. Policy.evaluate(req, rules, state) -> { decision, reason? } is pure — no IO, no async. Buyers, runners, and tests all share one decision function so allow/deny matrices stay consistent across deployments. Rules live as a JSON document referenced from the leash block on the registration JSON.

​
What policy covers

RulesV1 is intentionally small and inspectable:
Rule areaPurpose
BudgetDaily and per-call spend ceilings for the identity.
HostsAllow or deny lists for outbound destinations.
Price ceilingMaximum acceptable quoted price before settlement.
TriggersThe cron, interval, or webhook paths this identity is allowed to run from.
Stop conditionsConditions that stop the runner before more damage or spend accumulates.
Because policy is referenced from the identity registration, it becomes part of the trust surface. A service can ask not only “who is this agent?” but also “what is this agent allowed to do?”

​
Policy and receipts

Every policy decision should become observable. If policy allows a request and settlement succeeds, the receipt records decision: "allow". If policy blocks the request, the receipt records decision: "deny" and a reason. If policy allowed the request but settlement failed, the receipt records decision: "rejected". That closed loop — identity, policy, decision, receipt — is what lets Leash turn agent activity into reputation instead of relying on self-reported claims.