Jev Agent Skill Router

The project labels the numbers exploratory, reused-data results: 72 synthetic requests over 24 synthetic skills, and the routing questions were revised after inspecting an earlier full run. The original policy scored 64/72 overall and 52/60 on its initially untouched test split. The README states plainly that neither run establishes production accuracy or calibrated probabilities, and that typed output guarantees shape, not correctness.

It abstains a lot. 18 of 72 outcomes were reviews, 25.0%, against 12 for the lexical baseline, and the final run adds three unnecessary reviews plus one invalid_response whose raw body was not retained, so the precise bad field is unknown and no selective retry replaced it. Fourteen labels genuinely require review, and that volume is the cost of never misrouting.

Median 1,287 ms and p95 1,406 ms, with 287 API call attempts and 255,897 known input tokens, an estimated at least USD 0.0107 at advertised list rates rather than billing records. One failed call has unknown usage, which is why a complete final estimate would be misleading. The lexical baseline does the same 72 cases in 0.50 ms for no spend. Calls are bounded to four concurrent with no automatic retries.

Neither. It is an optional routing layer that returns route, no_skill or review, a three-valued specialist-needed judgement with its Noul, final Choice confidence and winner probability, and each call's candidates and distribution. It does not load skills or execute their instructions. The catalogue is entirely synthetic, and the 443-skill private catalogue that motivated the project is neither included nor used.

A route, no_skill or review outcome, a skill only when it routes, a three-valued specialist-needed judgement with its Noul probability, final Choice confidence and winner probability, every call's candidates and distribution, and measured latency plus API token usage.

Catalogues are sorted by name, split into deterministic batches and evaluated in parallel, with two candidates surviving each batch. Reduction repeats until a final Choice fits, with explicit no-skill and review options, and separate need, ambiguity and candidate-fit judgements gate the result.

Yes. jev-router route --baseline, evaluate --baseline-only, validate-results and report all run without calling the API, and report renders saved measurements. Live evaluation needs TYPESAFE_API_KEY in the environment, with no dotenv file or key flag.