A cookbook for AI that answers in probabilities, not prose

What it does

This is an independent, community-curated field guide of use cases, prompts, and starter patterns for Jev, TypeSafe AI’s “System One” model. It documents how to use the model’s three primitives—Choice, Score, and Noul—to get structured, probability-weighted answers that ordinary code can branch on for tasks like classification, routing, scoring, and guardrails.

The interesting bit

Jev occupies the awkward middle ground between brittle regex and expensive generative LLMs: you define the question and answer space in advance, and the model returns typed values with confidence scores. The README is admirably frank that this is not for writing prose or code—only for judgment calls that software needs to act on.

Key highlights

• Curates patterns for intent classification, semantic routing, reranking, fraud detection, and confidence-gated automation
• Three output primitives: Choice (pick one), Score (ordered rubric), and Noul (true/false estimate), each with probabilities
• Vendor-reported latency of 70–500 ms and input pricing at $0.042 per million tokens, with output tokens listed as free
• Accessible via direct API, Vercel AI Gateway, Cloudflare Workers AI, and official Python and JavaScript SDKs
• Explicitly community-maintained; warns that behavior, prices, and model aliases can change

Caveats

• Text-only modality: images, audio, and video are not supported
• English is the documented primary language; non-English workloads need representative testing before relying on calibrated thresholds
• The repository itself is not an official TypeSafe project

Verdict

Browse this if you are building agent workflows or safety layers and want LLM-assisted judgment without the parsing overhead. Skip it if you need generative output—the README itself suggests reaching for a general-purpose LLM instead.