What can System 1 answer without generating text?

Noul yes/no statements as P(true), choice among 2-16 options, and score on a 0-5 scale returning a distribution over the six levels plus an expected score, all in one prefill pass.

How is System 2 different from System 1?

System 2 is the untouched Qwen3.5-9B: the LoRA adapter is switched off inside peft.disable_adapter() and the model generates step-by-step reasoning, including the HumanEval completions.

What does the Escalate tab do?

It implements the card's confidence-gated pattern: System 1 answers in one pass, and when its top probability falls below your threshold the same unmerged weights hand the case to System 2.

Which measurements do the authors publish?

On 25,376 held-out decisions the mean KL to closed TypeSafe Jev 1.13 is about 0.019, ECE 0.0007, noul AUROC 0.994, and HumanEval 70.7% on the System 2 path; these are the authors' own numbers.

What are the documented failure modes?

System 1 mirrors its teacher's mistakes in multi-hop reasoning, arithmetic, dates and counting, reads English best, uses only the first 1,024 tokens of state, and handles at most 16 options.

Why is the first request after idle slow?

The 18 GB bf16 weights are streamed into VRAM by ZeroGPU on each request, so a call after an idle period is slower than the authors' ~90 ms median on a dedicated B200.

Is it safe for high-stakes decisions?

No. The card says it is not for high-stakes decisions and recommends confidence gating; for more than 16 options it points to autotrust/JEV-27B instead.

CommunityJev Demos & Experiments

JEV-9B Decision Demo

An interactive ZeroGPU demo of AutoTrust's JEV-9B: System 1 answers noul, choice, and score questions in one prefill pass with calibrated probabilities, while an Escalate tab hands low-confidence cases to the untouched Qwen3.5-9B base.

See highlights

Open huggingface.co

Format: Playground
Published by: Community
Author: AutoTrust AI
Last verified: Oct 2026

Tagscommunity playground demo gradio calibration

Works forJev for Classification Jev for Scoring Jev for Decision gating

Used inJev for Software & DevTools

Languages: Python
Platforms: Browser
Capabilities: Open sourceRuns locally

Highlights

System 1 returns a calibrated distribution over noul, choice (2-16 options), and score (0-5) decisions in a single prefill pass.
Authors report mean KL about 0.019 to closed TypeSafe Jev 1.13 over 25,376 held-out decisions, ECE 0.0007, and noul AUROC 0.994.
System 2 runs the untouched Qwen3.5-9B with the LoRA off, reaching HumanEval 70.7% with all 164 completions byte-identical to the base model.
The 18 GB bf16 weights load once at startup; ZeroGPU streams them per request, so the first call after idle is slower than the ~90 ms dedicated-B200 median.
The adapter is deliberately left unmerged so System 1 and System 2 live in one process; the decision head is a 24-slot fp32 layer.

Watch out

Apache-2.0, distilled from TypeSafe Jev 1.13's published output distributions rather than shared weights. The card says System 1 mirrors the teacher including its mistakes in multi-hop reasoning, arithmetic, dates and counting, reads English best, sees only the first 1,024 state tokens, and caps at 16 options. Not for high-stakes decisions; use the confidence gate.

Questions about this entry

Short answers drawn from the linked project or page.

Related terms

Glossary definitions related to this entry.

More Jev resources like this

JEV-27B-VL Demo

A ZeroGPU demo of AutoTrust's JEV-27B-VL: System 1 returns calibrated noul, choice, and score probabilities from one prefill pass over text and images, with System 2 on the untouched Qwen3.8-27B for escalation.