jevbooks

← All projects

Jevify

uspraveen/Jevify

Turn any open LLM into a calibrated System One decision model.

What it is

Jevify is an open-source benchmark, engine, and training recipe that gives any Hugging Face checkpoint the same typed-question interface as TypeSafe's Jev, for developers who need calibrated probability distributions instead of generated text.

How it uses Jev

Jevify reimplements Jev's System One contract: it reads a state and typed questions (choice, score, noul) and returns calibrated probabilities. The engine scores allowed answers by teacher-forcing against a shared prefix, then calibrates. A drop-in server exposes /v1/systemone and /v1/models, and the official typesafe-sdk works against it via TYPESAFE_BASE_URL.

Primitives:choicescorenoul

Technique worth stealing

Zero-initialized residual heads on the model's own scorer preserve its prior, avoiding the held-out-source regression of replacement heads.

Try it

Use the hosted API or run jevify-serve --model <checkpoint>; load_jevified("Praveenrajus/jevify-qwen3.5-4b").ask(state, questions).

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.89
  • Shows a System One patternno0.37
  • Handles uncertaintyno0.08
  • Measuredno0.10
  • Runnableyes0.71
  • Worth recommendingunclear0.47
  • Model replicayes0.89
  • Problem scopescore on a 0–2 scale1.09
  • About Jevyes0.98

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 21 Sept 2026.