stuntd

Checked, not fully reviewed. We ran its test suite and started its local mode, the one the README calls “local Jev without an API key”, against laya versions from 0.3.4 to 0.4.2. We did not run proxy mode in front of a paid provider or train a head. No Jev calls.

stuntd sits where your app’s provider base URL used to point. It speaks Jev’s /v1/systemone, the OpenAI Decisions API, Chat Completions and Anthropic Messages. It records the typed decisions passing through and distils each decision site into a small head on a frozen Laya encoder. Once a head is calibrated well enough, it answers that site locally and hands the uncertain cases back to the provider.

A fresh install did not start at 0.2.0. The README’s pip install "stuntd[train,jev]" allows any laya>=0.3.4, and pip picks 0.4.2. stuntd serve loads the checkpoint, fails its own warm-up question because the question’s instructions are empty, and exits. We bisected:

laya	stuntd serve
0.3.4, 0.3.20, 0.3.21	starts
0.3.24 to 0.4.2	exits after the warm-up

Six tests in one file fail from 0.3.20 on, for a related reason: a test double’s tokenizer does not accept an argument newer laya passes. Its CI is green because it never installs laya. We reported it.

Fixed in 0.2.1, the same day. The warm-up now asks a real question, laya is capped below 0.5, and CI installs the newest laya. On 13 October a fresh pip install "stuntd[train,jev]" picked stuntd 0.2.1 with laya 0.4.6, and the README’s three-command quickstart served local Jev. The official SDK got its answers, with model set to stuntd and the X-Stuntd header saying mode=local.

Once running, it is honest about who answered. With laya 0.3.20, the official Python SDK worked against local mode unchanged. The response’s model was stuntd, not jev-latest. The X-Stuntd header said jev; mode=local; questions=1; live=0; zeroshot=1, and /v1/models listed only stuntd. The README’s rule is “every answer says how it was served, so a silent fallback cannot hide”, and on the Jev path that holds in the body as well as the header. On the OpenAI-compatible path, its code builds a locally served answer with the model name your request asked for, so there the header is the only signal.

Zero-shot Laya’s limits come through. For “I was charged twice.”, the question “Is it false that this is a billing problem?” came back 0.803 yes. See our Laya review.

Apache-2.0, 70 stars.

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

Review updated Oct 13, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.