agent-jev
AgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any unstructured state (diffs, traces, logs) and structured questions, get calibrated
What is agent-jev?
AgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any unstructured state (diffs, traces, logs) and structured questions, get calibrated probability distributions back in one ~50ms forward pass. Zero output-token decoding.
Overview of agent-jev
What it is
AgentJev-0.6B is a fast 'System One' decision model for AI agents. It accepts unstructured state (diffs, traces, logs) and structured questions, and returns calibrated probability distributions in a single forward pass of approximately 50ms. No output tokens are decoded; the model does not generate text.
Key features
- Zero output-token decoding: probabilities are returned directly from hidden states.
- Three question primitives: Boolean, Choice (2–255 options), and Score (2–10 levels). Each returns the full distribution, not just the winner.
- State is encoded once for all questions; a tree encoder merges token prefixes, reducing backbone tokens per case from 4,074 to 443 and median latency from 278 ms to 44 ms on an RTX 5060 (measured on a 400-case test split).
- Context window of 2,048 tokens; inputs exceeding this are refused.
- Based on Qwen3-0.6B with the language-model head removed and a permutation-equivariant candidate head added.
- On the Typed Decisions test split (400 cases, 2,000 questions), achieves 79.25% top-1 accuracy, with soft cross-entropy 0.8494 and Brier score 0.0448.
Use cases
Place AgentJev where an agent loop needs a gate, a route, or a score. Examples: determining whether tests passed, selecting the next tool, assessing if a command is safe to run. It handles quick decisions, leaving text generation to larger models.
Getting started
See the repository README for installation and quick-start commands.
License and pricing
Licensed under Apache-2.0. Open source.
Based on the project's public README and repository metadata.