README.md

metadata

title: JEV-9B Decision API
emoji: ⚖️
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 6.28.0
app_file: app.py
python_version: '3.12'
startup_duration_timeout: 1h
pinned: false

JEV-9B Decision API

A minimal, machine-friendly wrapper over System 1 of autotrust/JEV-9B — AutoTrust AI's open decision model (Apache-2.0), built with the Blocks-of-Experts recipe on a frozen Qwen3.5-9B backbone.

System 1 answers a typed question in a single prefill pass (nothing is generated) and returns a calibrated probability distribution:

• noul — "Is this true?" → [P(false), P(true)]
• choice — which of 2–16 options → one probability per option
• score — where on an ordered 0–5 scale → distribution over six levels

API

A single Gradio endpoint, /decide, taking kind, state, question, options (comma-separated, choice only) and returning JSON:

{"kind":"choice","distribution":{"a":0.53,"b":0.33,"c":0.14},"best":"a",
 "confidence":0.53,"state_tokens":18,"truncated":false}

This Space exists to give local agents (Hermes) one clean JSON call, rather than the multi-question HTML UI of the upstream demo. Inference core (jev_core.py) is the model authors' documented transformers + peft path, reused verbatim.

Provenance & limits

jev_core.py is reused unmodified from the authors' official demo Space (autotrust/jev-9b-decision-demo), which is Apache-2.0. Model card caveats still apply: System 1 mirrors its teacher including its mistakes (multi-hop reasoning, arithmetic, dates, counting, adversarial inputs), reads English best, uses the first 1,024 tokens of state, and is not for high-stakes decisions — use confidence gating and route low-confidence decisions elsewhere. Runs on ZeroGPU (zero-a10g).