---
title: JEV-9B — typed decisions
emoji: ⚖️
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 6.28.0
app_file: app.py
short_description: Calibrated typed decisions + System 2 reasoning
python_version: "3.12"
startup_duration_timeout: 1h
pinned: false
---

# JEV-9B — typed decisions in one forward pass

An interactive demo for [`autotrust/JEV-9B`](https://huggingface.co/autotrust/JEV-9B) — AutoTrust's
first integrated **System 1 + System 2** open model, built with the Blocks-of-Experts recipe on a
frozen **Qwen3.5-9B** backbone (Apache-2.0).

**System 1** answers typed questions in a *single prefill pass* — nothing is generated — and
returns a **calibrated probability distribution**:

| kind | question | returns |
|---|---|---|
| `noul` | "Is this statement true?" | `[P(false), P(true)]` |
| `choice` | "Which of these 2–16 options?" | one probability per option |
| `score` | "Where on the ordered 0–5 scale?" | distribution over the six levels + expected score |

**System 2** is the untouched base model: the LoRA adapter is switched off and Qwen3.5-9B
generates step-by-step reasoning (thinking mode). The **Escalate** tab implements the model
card's confidence-gated pattern — System 1 answers in one pass, and when its top probability is
below your threshold the *same weights* hand over to System 2.

## What to expect

Per the authors' published measurements: mean KL ≈ 0.019 to the closed TypeSafe Jev 1.13 on
25,376 held-out decisions (indistinguishable at that resolution), ECE 0.0007, `noul` AUROC 0.994,
and HumanEval 70.7% on the System 2 path — all 164 completions byte-identical to the base model.
System 1 mirrors the teacher *including its mistakes* (multi-hop reasoning, arithmetic, dates,
counting), reads English best, uses the first 1,024 tokens of state, and handles at most 16
options — for longer lists the authors recommend
[`autotrust/JEV-27B`](https://huggingface.co/autotrust/JEV-27B). Not for high-stakes decisions;
use confidence gating.

## Implementation notes

- Inference follows the model authors' documented `transformers` + `peft` path 1:1: backbone with
  the LoRA adapter on, the 24-slot fp32 decision head applied to the last token's post-norm
  hidden state, divided by the per-kind calibration temperature (noul 1.002 · choice 0.984 ·
  score 1.012, from `calibration.json`). System 2 runs with the adapter disabled inside
  `peft.disable_adapter()`. The adapter is deliberately left unmerged so both systems live in
  one process.
- Weights (18 GB bf16) load once at startup; ZeroGPU streams them into VRAM on each request, so
  the first call after an idle period is slower than the ~90 ms median the authors measure on a
  dedicated B200.
- The demo runs on ZeroGPU (`zero-a10g`); each decision costs one prefill pass, and System 2
  generations are capped by the token-budget slider.

## Example data

The `gr.Examples` rows come from the model authors' own published material (Apache-2.0): the
quickstart cases in the model card and the Hacker News front-page / community cases in
`reports/realworld_9b.json`, with expected answers published alongside them.

## Credits

Model: [AutoTrust AI](https://huggingface.co/autotrust) — Apache-2.0, distilled from the
published output distributions of TypeSafe Jev 1.13 (an independent student that shares no
weights, code or affiliation with TypeSafe AI). System 2 path:
[`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B).