jevbooks

← All projects

openJev-verdict-2.0

Heman10x-NGU/openJev-verdict-2.0 · Homepage

Calibrated 151M non-autoregressive decision engine for typed software workflows.

Evaluation & Benchmarking29%Runner-up: Calibration & Research

What it is

openJev-verdict-2.0 is a 149.6M parameter decision model that evaluates typed decision schemas (Choice, Score, Noul) in non-autoregressive forward passes with calibrated confidence. It targets enterprise financial, security, customer support, and agent trace observability workflows.

How it uses Jev

The model takes a state and typed questions (Choice, Score, Noul) and returns calibrated probabilities. It uses dual-channel calibration: a distribution head for soft labels and a confidence head for correctness. Deterministic software retains control over thresholds and business rules, using the model for bounded semantic classification.

Primitives:choicescorenoul

Technique worth stealing

Symmetric Permutation-KL penalty between option-shuffled batches reduces prompt-order bias.

Try it

Run `cd webgpu-demo && python3 -m http.server 8080` and open http://localhost:8080 in Chrome or Edge.

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-centricyes0.83
  • Shows a System One patternno0.25
  • Handles uncertaintyno0.05
  • Measuredyes0.94
  • Runnableno0.18
  • Worth recommendingno0.24
  • Model replicayes0.93
  • Problem scopescore on a 0–2 scale1.35
  • About Jevyes0.91

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