Overview & result
Jeved is a SillyTavern extension. It sends chat messages to Jev with questions that you write; a rule reads the answers and acts by adding one instruction to the prompt, rerolling the reply, changing a list or running an STscript. When no rule matches it adds nothing. The README's examples include pacing, steering, rerolls, narrator voice and inventory state.
From the original post
SillyTavern extension that measures each reply and instructs the narrator only when a rule matches.
How Jev fits in the loop
- You write sensors: a Score, Choice or Noul question about recent chat messages.
- Jeved sends the messages and questions to Jev before or after a reply.
- A rule reads the answers and acts: it adds one instruction to the prompt, rerolls the reply, changes a list or runs an STscript.
- When no rule matches, nothing is added to the prompt.
How to reproduce
- Inspect the original showcase and source material at https://github.com/mossyfield/ST-jeved
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
The narrator is only instructed when a measured condition is met, instead of carrying standing instructions in every prompt.
Limitations
- Needs SillyTavern 1.18.0 or later and an API key.
- A sensor that runs before the reply adds one call per message; Jeved waits at most 3 seconds for it.
- The README gives no evaluation of how accurate the measurements are.