Jev Model Router — gualican

For every prompt it makes one call to TypeSafe asking two typed questions: tier as a Choice among simple, moderate, complex and frontier, and stakes as a Score from 0 to 2 estimating how costly a wrong answer would be. The resolved tier is returned to your application.

Two rules in router.py: a tier classification below CONFIDENCE_FALLBACK_THRESHOLD falls back to moderate, and stakes above STAKES_OVERRIDE_THRESHOLD escalate the request one tier up even if confidence was fine. The thresholds live in config.py and the README says to tune them.

No. It never calls OpenAI or any other downstream provider; it resolves a tier, maps it to a model in config.py, and returns the decision. Your application makes the actual model request, so the router stays provider-neutral in its call path.

After pip install -e ., use jev-router route 'Write a haiku about databases' to classify and print the selected model, jev-router classify to print the raw Jev classification, or jev-router serve --reload to start the HTTP server. Without installing, python -m jev_model_router.cli route works the same way.

Run jev-router serve --reload and POST JSON to the local endpoint, for example curl -X POST localhost:8000/route with a prompt field, which is equivalent to the CLI route command. The README documents this as the server mode of the same classifier.

No — the mapping and thresholds in config.py are documented as starting points to adjust from observed results. Note also that the GitHub description names Claude tiers while the README maps tiers to OpenAI model names; check the config and the code before relying on either naming.

pytest runs unit tests with mocked TypeSafe calls, so no API key is needed for the test suite. scripts/manual_eval.py runs real Jev classifications against representative prompts when you want to see live behaviour, and the README recommends calibrating thresholds from such observations.