Inside JEV: the $0.042/M-token fast-judgment model, black-box tested and dissected

AGI Hunt · wechat · 2026-09-20

A deep dive into TypeSafe AI's JEV, founded by ex-OpenAI researcher Diogo Almeida with a $40M seed round: 140k waitlist signups in 36 hours and a day atop Hacker News. What it is: a decision-oriented model that takes a state plus question definitions and returns parallel typed answers with probabilities — not autoregressive text — at $0.042/M input tokens, 70–500ms latency, suited to high-frequency classification/routing rather than deep reasoning.

Black-box findings (Archer Hume): Choice/Score confidence is a distribution statistic, not an independent correctness probability; Noul's pro/contra answers can sum above 1; batching 13 questions is 10x faster than sequential with identical answers; adding an option shifts existing options' log-odds (+0.38), suggesting contextual interaction.

Architecture & replicas: officially a pretrained Transformer post-trained via RLHF, RLVR, and the novel RLCD optimizing calibrated decisions — recipe undisclosed; "one prefill + readout head" is plausible but unconfirmed. Logit-based zero-training replicas (with temperature scaling) achieve 5–7x speedups, but training-time calibration and post-hoc scaling aren't equivalent, and masking invalid tokens ≠ task understanding.

Original post →

Inside JEV: the $0.042/M-token fast-judgment model, black-box tested and dissected

More from coding & agent