Module policy 

AI routing policies (DW-085): within-request escalation and latency-vs-cost selection. Composed over DW-076 routing.

Two policy kinds:

• FallbackChain: calls an external classifier service to estimate prompt complexity. Simple prompts (score < threshold) route to the cheap model; complex prompts (score >= threshold) escalate to the costlier model. On classifier error, fails open to the cheap model (the safe default).
• LatencyCost: static config-based selection. Candidates are sorted at compile time by the configured preference (cost, latency, or balanced); the policy returns the best candidate deterministically (no runtime metrics needed).

A policy composes over DW-076 routing: the candidate aliases it names (cheap, escalate_to, or candidates[].model) are themselves plain chain/canary aliases. The policy resolves each alias to its PRIMARY RouteTarget at compile time (the first chain entry or the first canary version), so the runtime evaluation returns a flat candidate list the dataplane walks with the same failover loop as a plain chain.

Dependency direction

ai depends on config only. The classifier HTTP call reuses the same hyper_util client pattern as the DW-083 semantic cache (no new dependencies).

Structs

LiveLatencyTracker

AI-15 (#203): Live latency tracker for LatencyCost policies. Maintains a rolling window of latency observations per model alias. Thread-safe (wrapped in Arc).

Enums

CompiledRoutingPolicy

A compiled routing policy (DW-085). Built at AiRuntime compile time from the config block; immutable once built. The evaluate method returns the candidate list to walk (the same shape as AiRuntime::route) plus a PolicyDecision the dataplane records as a metric.

PolicyDecision

The decision a routing policy made for one request (DW-085). The dataplane records this as a metric (the ai domain cannot import observability — the dependency direction forbids it — so the decision travels back to the caller, which records it).