README
JEV-9B — typed decisions in one forward pass
An interactive demo for 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. Not for high-stakes decisions;
use confidence gating.
Implementation notes
- Inference follows the model authors' documented
transformers+peftpath 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, fromcalibration.json). System 2 runs with the adapter disabled insidepeft.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 — 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.
Similar repos
README
JEV-9B — typed decisions in one forward pass
An interactive demo for 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. Not for high-stakes decisions;
use confidence gating.
Implementation notes
- Inference follows the model authors' documented
transformers+peftpath 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, fromcalibration.json). System 2 runs with the adapter disabled insidepeft.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 — 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.