Klyr does not treat model output as authorization.
Model output is not authorization
An AI model may propose a structured claim, rank candidates, or extract resolution conditions. It never declares a market verified or approves execution.
Deterministic policy enforcement
Risk policy - position limits, liquidity thresholds, duplicate protection - is evaluated by fixed, auditable logic, not model judgment.
Fail-closed behavior
When market data is unavailable, resolution rules are ambiguous, or a required check cannot complete, Klyr responds with BLOCK or ESCALATE - never a silent guess.
Explicit ambiguity states
A vague deadline or unclear threshold is surfaced as an ambiguity, not resolved by assumption.
Execution idempotency
Every execution requires an idempotency key. A replayed request returns the original decision receipt rather than executing twice.
Bounded requests and timeouts
External calls are time-bounded. A timeout produces a typed, graceful failure rather than a hang.
Verification snapshots and receipts
Every decision preserves what Klyr actually knew at the time: the claim, the market, the resolution rules, the state, the policy result, and a decision hash.