jevbooks

← All projects

jevcal

abhixhek/jevcal

Calibrate, threshold, and drift-check typed decision models against an LLM teacher.

What it is

jevcal measures a typed decision model on your labeled data, picks per-question confidence thresholds to meet your accuracy target, reports escalation traffic, and fails CI on drift. Built for Jev and any model returning answers with probabilities.

How it uses Jev

Jev answers typed questions (Choice, Score, Noul) on your state. jevcal runs the model over your dataset, saves predictions, then compiles thresholds from confidence measures like top_prob, margin, entropy, or the API's confidence. The lock file and report show accepted accuracy, handled share, ECE, and cost split.

Primitives:choicescorenoul

Technique worth stealing

Pick thresholds on one half of your data and verify on the other half; optionally require the 95% lower confidence bound to clear the target.

Try it

pip install "git+https://github.com/abhixhek/jevcal" then jevcal demo

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-centricunclear0.58
  • Shows a System One patternunclear0.56
  • Handles uncertaintyyes0.89
  • Measuredyes0.96
  • Runnableyes0.67
  • Worth recommendingyes0.61
  • Model replicano0.06
  • Problem scopescore on a 0–2 scale1.48
  • About Jevyes0.96

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.