AgentJev trades token generation for a single forward pass over typed decisions

AgentJev-0.6B is a routing and gating model that produces zero decoded tokens. You hand it a state — a diff, a stack trace, a ticket thread — plus typed questions, and one forward pass returns a probability distribution over the options you supplied. No generation, no JSON to parse, no output vocabulary step.

The architecture strips the language-model head from Qwen3-0.6B and replaces it with a small permutation-equivariant candidate head. Each candidate is read at its last token; softmax is per question. Because the order of options cannot smuggle in a ranking, the model is forced to judge on meaning alone. Three primitives cover the decision space: boolean (a proposition with optional criteria for each branch), choice (up to 255 options, returning the full distribution plus a margin between first and second), and score (2–10 ordered levels, with an expected score computed as a probability-weighted sum).

On the Typed Decisions benchmark — 400 cases, 2,000 questions across four workflows — the published checkpoint hits 79.25% top-1 accuracy, against 77.00% for Laya, the prior specialist. The bootstrap 95% interval on that gap is +0.65 to +3.90 points. The README is careful about what this means: accuracy is agreement with a teacher argmax, not a measured rate of merged pull requests or resolved incidents. On security incidents and agent-trace observability, Laya still leads by about one point each.

The latency story is conditional. On short five-question cases, Laya's smaller 421M ModernBERT encoder is roughly 20 ms faster at the median. AgentJev's advantage appears under wide candidate sets: a 64-option choice question plus one boolean, totalling 33,547 path tokens, runs in 298.91 ms with shared-prefix KV caching versus a 500–600 ms band for Laya, which has no causal prefix seam and must re-encode the state for every candidate. The README puts the token reduction at 92.4%, from 33,547 to 2,551 unshared tokens. The context ceiling is 2,048 tokens; over-length input is refused rather than silently truncated.

The practical hook is agentjev_hook.py, a Claude Code PreToolUse command that posts Bash, Write, and Edit payloads to the local server, asks a boolean and a four-level risk score, and blocks only when the score hits level 3 and the boolean says the action is unsafe. If the server is unreachable, the hook exits 0 and the tool proceeds — a sensible fail-open default for a reflex layer that is not meant to be the last word.