@velum-labs/routekit-eval-contracts
v1.3.2
Published
Versioned RouteKit evaluation run, result, evidence, and policy contracts.
Readme
@velum-labs/routekit-eval-contracts
Versioned evaluation run, result, evidence, and policy contracts encoded as Effect Schema. Candidate and judge calls always name explicit model IDs; evaluation never uses the auto-router.
Automatic routing
The sole automatic-routing protocol keeps request understanding separate from model selection:
- A reviewed basis defines five to ten semantic routing dimensions.
- The classifier returns one weight for every dimension plus
unknownWeight. - A published snapshot supplies a complete candidate-model by dimension evidence matrix.
- A deterministic objective ranks eligible models from that vector and evidence.
The classifier contracts contain no model IDs, prices, evidence, winners, or fallbacks. Use the exported assertion functions after schema decoding to enforce relationships that span multiple fields or documents:
assertRoutingBasisassertDecompositionInputassertDecompositionResultassertRequestDecompositionassertRoutingObjectivePolicyassertPublishedRoutingActivationassertAutoRoutingDecision
These checks reject incomplete or duplicate vectors and evidence cells, definition/evidence digest mismatches, invalid model IDs, non-normalized objectives, and misleading cost averages when any calls are unpriced.
