agent-jev

<p align="center"> <img src="assets/agentjev_reflex_demo.gif" alt="AgentJev returns a distribution in one forward pass. A writer model is still decoding." width="100%"> </p> <p align="center"> <a href="README_zh.md"><img alt="简体中文" src="https://img.shields.io/badge/lang-简体中文-111111?style=flat-square"></a> <a href="https://github.com/malevrigns/agent-jev"><img alt="Code" src="https://img.shields.io/badge/code-GitHub-111111?style=flat-square"></a> <a href="https://github.com/malevrigns/agent-jev/blob/main/LICENSE"><img alt="Apache-2.0" src="https://img.shields.io/badge/license-Apache_2.0-111111?style=flat-square"></a> <a href="https://huggingface.co/Qwen/Qwen3-0.6B"><img alt="Qwen3-0.6B" src="https://img.shields.io/badge/backbone-Qwen3--0.6B-111111?style=flat-square"></a> <img alt="FP32" src="https://img.shields.io/badge/weights-FP32-C6A36A?style=flat-square"> <img alt="Zero decoded tokens" src="https://img.shields.io/badge/decoded_tokens-0-C6A36A?style=flat-square"> </p> <p align="center"> <b>AgentJev-0.6B</b><br> State in. A distribution out. Nothing decoded.<br> <a href="#scores">Scores</a> &nbsp;&nbsp;·&nbsp;&nbsp; <a href="#use-it">Use it</a> &nbsp;&nbsp;·&nbsp;&nbsp; <a href="#one-forward-pass">Forward pass</a> &nbsp;&nbsp;·&nbsp;&nbsp; <a href="#how-to-read-the-scores">Reading the scores</a> </p>

This file is the coding-completion checkpoint. It reads the task, the code, and the checks that were shown, then returns a probability that the work is actually finished.

The backbone is Qwen3-0.6B with the language-model head removed. A permutation-equivariant candidate head scores the options you supply, and the order of those options does not become a ranking. Softmax is per question. Boolean, Choice, and Score each come back as the full distribution. Code and client: malevrigns/agent-jev.

Scores

Supervised training and RLCD use executed coding pairs. Both sides pass the same public checks, and the label comes from the checks that were not shown. The frozen development set selected RLCD step 12. Selection was locked before the test split was opened.

Invoice processing is agreement with the public teacher argmax on that workflow.

Use it

model.safetensors is FP32. It is the full module: backbone plus candidate head. The server wants a torch checkpoint, so wrap the file once. The process binds 127.0.0.1. The workbench is http://127.0.0.1:8149/.

git clone https://github.com/malevrigns/agent-jev.git
cd agent-jev
python -m venv .venv
pip install -r requirements.txt
pip install huggingface_hub safetensors

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import torch

src = hf_hub_download("aimeigaoshou/agent-jev", "model.safetensors")
torch.save({"state_dict": load_file(src)}, "agentjev.pt")
hf_hub_download("aimeigaoshou/agent-jev", "temperatures.json", local_dir=".")

python -m jev_service.server \
  --checkpoint agentjev.pt \
  --model-path Qwen/Qwen3-0.6B \
  --temperatures temperatures.json \
  --port 8149

from agentjev_client import AgentJev

jev = AgentJev("http://127.0.0.1:8149")

done = jev.decide_boolean(
    {
        "task": "Fix the expired-token failure in UserAuthService.verifyToken",
        "test_output": "Tests run: 14, Failures: 1 — test_expired_token",
    },
    "Does this implementation meet the stated requirements?",
    criteria={
        "true": "The requirement is met.",
        "false": "A requirement is still unmet.",
    },
)

done["prob_true"] is the probability the work is finished. done["decision"] is the argmax. Choice and score go through decide_choice and score in the same client. Question ids are echoed to the caller. The model sees the question text and the option text.

One forward pass

<p align="center"> <img src="assets/agentjev_shared_prefix.gif" alt="Shared-prefix cache scores many candidates from one encoding of the state." width="100%"> </p>

Candidates of the same question reuse one prefix. On a fixed load — 64 choice options plus one boolean, 33,547 path tokens — the unshared median was 609.65 ms and the shared-prefix path was 298.91 ms. Backbone token ops went from 33,547 to 2,551. The largest probability gap was 0.000508, and the chosen option stayed the same. That figure is this load, after warmup.

The server ceiling is 2,048 tokens. Input past that ceiling is refused, so a long diff is an error instead of a silent crop of the question.

How to read the scores

Coding accuracy is agreement with the executed label on the coding-completion set. Invoice processing is agreement with the teacher argmax. A win on either row is a score on that set. It is a separate question from whether a pull request merged or an invoice was paid.

The typed-decisions write-up lives in the code repository. The tensors in this repository are the coding-completion checkpoint above.

Citation

@misc{agentjev2026,
  title        = {AgentJev: A 0.6B System One Decision Model},
  author       = {malevrigns},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/aimeigaoshou/agent-jev}},
  license      = {Apache-2.0}
}