Jeved

How we reviewed this: we ran its test suite in Docker on Node 24 and drove its classifier module directly against a local stand-in to see what a judgment request carries. We did not install SillyTavern, so the interface and the in-session behaviour are read rather than run — the rules engine below is described from its code and its tests.

What it does

Jeved is a SillyTavern extension that asks questions about the conversation as it goes. You write the questions — “has the scene been calm for several replies?”, “is the character contradicting something established?” — and each generated reply gets scored. A rule over those scores can then add one line to the prompt, reroll the reply, or run an STscript.

That is a genuinely good fit for the model. Pacing, continuity and tone are exactly the things a long roleplay prompt asks for and does not get, because instructions decay while a check after each reply does not.

What it sends

We called the classifier against a stand-in with a two-message chat:

{"model": "jev-1.13.0",
 "state": {"chat": [{"name": "Alice", "mes": "We have been walking this corridor…"},
                    {"name": "You",   "mes": "I look for a door."}]},
 "questions": {"pacing": {"type": "noul", …}}}

The chat text goes out, once per judged reply, to whichever of three hosts you configure:

route	endpoint
OpenRouter	openrouter.ai/api/alpha/decisions
NanoGPT	nano-gpt.com/api/v1/decisions
TypeSafe	/proxy/https://api.typesafe.ai/v1/systemone

The third one is worth explaining: it goes through SillyTavern’s own CORS proxy, which needs enableCorsProxy: true in config.yaml — the extension’s hint says so. That is the route for people who already have a TypeSafe key; the other two let you reuse an account you probably already have for the generative side.

Retries are limited to 429 and 529 with a fixed delay, which is the right pair to retry and a short enough list not to hide real errors.

The test suite is the surprise

621 tests, all passing, on a repository with three commits, for a SillyTavern extension. That is more coverage than most of the paid tooling we review. They cover the rule engine’s conditions and consecutive-match counting, the sensor types, the macro and preset handling, the context grouping, and the classifier’s error paths.

For an extension whose whole job is “watch every reply and sometimes intervene”, that coverage is the difference between a fun weekend and something you leave running in a long story.

Verdict

If you use SillyTavern and have ever wanted the narrator to notice that nothing has happened for six replies, this is the right shape for it: your questions, your thresholds, your action, checked after each reply rather than hoped for in a system prompt.

Know that your chat text leaves the machine on every judged reply, pick the host deliberately, and read the rules you write as carefully as you would a prompt — a rule that rerolls on a low score will happily reroll forever if the score never moves.

For the same idea outside roleplay, see jevmeter for transcripts and compact-adviser for agent sessions.

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

Review updated Sep 21, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.