Reproducible routing evidence · MIT + CC BY 4.0

Multi-provider route-admission fixture

See why a deployment is allowed to receive an LLM task before any provider call happens. The bundle evaluates eleven synthetic deployments against capability, tenant, region, evaluation, context, deadline, cost, and health rules. Every rejection remains visible in the resulting manifest.

This run proves static admission behavior, not provider quality or runtime failover. Its deployments and prices are fixed test inputs, not live vendor data.
Task policy filtering eleven synthetic LLM deployments through eight admission gates before selecting an eligible route

Files

Admission before ranking

A cheap or fast deployment cannot outrank a failed tenant, region, capability, or evaluation rule.

Deterministic selection

Eligible routes use declared priority and stable deployment ID, so registry input order cannot change the result.

Inspectable rejection

The manifest retains every rejected deployment, reason, policy version, and registry version needed for incident review.

Verified scope

Nineteen tests cover capability, tenant, region, evaluation, context, deadline, cost, deployment state, stable ordering, input immutability, version evidence, duplicate IDs, stored JSON, and the generated SVG. The reviewed run admitted two of eleven candidates and selected the declared priority-10 route.