tiershift

How we reviewed this: we ran both implementations’ test suites in Docker offline — 118 TypeScript tests and 85 Python tests passing — and recomputed its published benchmark from the results file in the repository. We made no Jev calls.

What it does

tiershift, by Vatsal Patel, routes each request to a model tier. One request asks eleven signals — difficulty, stakes, whether reasoning is needed, safety sensitivity, domain, expected output length and more — and a YAML policy turns those into a tier, with a printed reason chain: default, then which rule overrode it and why.

It ships a TypeScript implementation, a Python one, and an OpenAI-compatible proxy.

The benchmark checks out

Its README publishes a comparison across four routing arms on 120 prompts. The per-request results are committed, so we recomputed the table ourselves: quality and cost per thousand requests matched the published figures cell for cell.

That is rare. Most projects publish a number; this one publishes the rows behind it.

What to know

The state carries the system prompt (truncated), the last user message (up to 4,000 characters), the previous assistant message and the last tool result. The decision log keeps signals, tier, model, cost and reason — we checked a real log line, and it contains no message text, as the README claims.

One inconsistency: the headline latency figure differs between the README and the two package manifests.

Verdict

The most check-able small project in this round: two implementations, both tested, a policy you can read, and a benchmark whose arithmetic survives an independent recomputation.

For other routers, see jev-router and jev-gateway.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 20, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.