Policy evaluation

Candidates observe matching events without denying access. Routes use exact path-template equality or *, not arbitrary patterns. Services and opaque subjects are host-provided identifiers.

Tenant isolation

tenant_isolation checks that actor_tenant and resource_tenant are present and equal. Missing/mismatched context is a violation; an active policy can deny the attempt. Derive identity and ownership from trusted server state, not arbitrary request parameters. Matching tenants do not establish all permissions.

Rate limiting

rate_limit counts attempts per (service, route, subject), including denied attempts. threshold ranges from 1 to 100,000; window_seconds from 1 to 3,600. Counters use server receipt times and retained events. Pruning a recent window changes counts. This is not a globally distributed limiter.

Response monitoring

error_burst monitors response status ≥400; latency monitors observed latency above its threshold. These concern completed responses and cannot block pre-operation requests. Sending the same attempt to /events and /decisions duplicates counts rather than updating a response.

Historical simulation

POST /v1/policies/{id}/simulate evaluates a retained snapshot and returns its watermark, matched/affected counts, label breakdown and up to 100 affected event IDs. Replay errors above 100,000 retained events/account. Later events are absent; no real business operation is repeated.

expected labels (legitimate, abusive, unknown) come from the host. False-positive metrics describe that retained labeled sample only. Incorrect labels, small samples and old history do not establish future safety.

Activation, expiry and revocation

Activate an enforceable kind using a matching simulation less than 15 minutes old, relevant events and acknowledge_impact: true. Manual activation requires operator review; legitimate impact is not automatically prohibited. AI administration separately checks configured replay criteria against fresh history.

One active policy's denial is sufficient. ttl_seconds is 1–86,400; expiry is checked during evaluation without a separate scheduler. Revocation stops enforcement. Recreate/review a candidate before reactivation. allowed: true only means no active Qinzora policy denied the attempt; retain host authorization.