MEDIA PREVIEW
Overview & result
A Python skill router uses Jev Choice and Noul judgments to return route, no_skill or review for a described skill catalogue. In an exploratory synthetic evaluation, the author reports 68/72 exact outcomes versus 51/72 for a lexical baseline, with 1,287 ms median request latency; it does not execute the selected skill.
From the original post
A small Python CLI and library accepts a request and a catalogue of skill names with public descriptions. It returns route, no_skill or review, with a selected skill only for route.
How Jev fits in the loop
- Sort the public skill catalogue and split large sets into deterministic batches.
- Use Jev Choice calls to retain top candidates from each batch, then ask final Choice and Noul questions.
- Apply thresholds in Python to return a skill, no-skill result or human-review abstention.
How to reproduce
- Install the Python CLI and replay the repository's published baseline and final-result files.
- Supply a catalogue of skill names and public descriptions, then inspect route, no_skill or review output.
- Repeat a live run with a TypeSafe key and compare all outcomes, failures and API latency rather than only successful routes.
Why this build matters
Provides a reproducible skill-routing experiment with all 72 synthetic cases and explicit error and abstention counts, while separating routing from skill execution.
Limitations
- The published 72 requests and 24 skills are synthetic; the private 443-skill catalogue was not used in the public benchmark.
- The final routing policy was revised after an earlier full run, so 94.4% is exploratory rather than production accuracy.
- The library routes or abstains; it does not load or execute a skill.