What I Use Jev For

Jev is TypeSafe’s System One model that doesn’t generate prose but answers typed questions (yes/no, pick-one, scored) with probabilities your code can branch on. Priced at $42 per billion input tokens, it’s integrated as a judge() interface in Oh My Pi: you send a state and questions and get numeric likelihoods (e.g., "The sky is green" → 0.05). Over two weeks Jev ran 629 judgments across five illustrative workflows for a total cost under $0.09, with most calls extremely cheap and sub-second median latency. Its semantic-search role drives a pipeline that pre-ranks files, judges the top 128 by filename, excerpts from the best 20, then full passages, returning scores the agent uses to choose actions.

In practice Jev powered automatic code search (604 calls, $0.0884) to find helpers and routes (real examples: an email helper scored 0.94; a /metrics route and related files ranked 0.89-0.23), supported quick design synthesis from multiple files, ran multi-question judge cells (five judgments at once with concrete probability outputs), and auto-selected thinking effort for subagents (19 calls, cost ~ $0.0004; decisions ~ $0.00002 each). The result: cheap, precise probabilistic judgments that let agents decide when something is a strong match, when to dig deeper, and how much effort to expend.