Decide
in one
pass.

Judgment models for agents. One forward pass, calibrated probabilities over the options you define. No generated text, nothing to parse.

$ pytest -x · tool outputillustration
q · does this line matter?kept 0 / 0
Generation vs. judgment

A chat model writes its answer. A judgment model reads it off.

LLM · generating0.0s

Judgment model · one pass0.0s

refund
0.82
logistics
0.13
security
0.05
Illustration of the mechanism; timings not measured. Benchmarks will be published with the model.
Current work
■ Training

An open judgment model

Open weights and a public recipe: which objective yields calibrated probabilities, base or instruct start, dense or MoE.

■ Running

Judgment in the agent loop

A coding-agent harness that filters every tool output through a judgment model. Filter on vs. off, measured on cost, tokens and pass rate.

■ Ongoing

Evaluation beyond accuracy

Same questions for every model: accuracy, agreement when options are reordered, NLL, Brier score, calibration error.