The 16 Jev Design Patterns

Four areas: what goes to the model versus code (Viewfinder, Menu, Question bank, Ghostwriter), how one request is shaped (Questionnaire, Prefetch, Blind review, Funnel), how a probability becomes an action (Traffic light, Grey zone, Thermostat, Scorecard), and how you know it holds (Inspection, Fail open/fail closed, Flight recorder, Check-up).

Yes. The Check-up pattern, measured calibration, covers grading questions with Brier scores over time. The validation group also includes contract validation for malformed answers, a failure-mode knob for choosing fail-open or fail-closed behavior, and journal-and-replay for recording probabilities rather than actions.

Traffic light is tiered thresholds in a probability-to-action group: red stops, amber looks, green goes. It sits next to the Grey zone pattern, which holds that 0.5 is not a medium reading but an unknown, and Thermostat hysteresis, which enters at one level and exits at another, on at 20 and off at 22, so the action does not flap on a single borderline answer.

Viewfinder is state minimalism, framing only what is being judged rather than passing a whole document. Menu is candidate enumeration: code orders what is on the menu, then asks a noul, noul or score per candidate, so the model chooses from a fixed list instead of composing an answer.

Each listing is a problem you will run into, the way out, and the projects that took it, illustrated with small ASCII diagrams. Alongside the 16 patterns sits a gallery of 1,354 projects built on Jev, and the site says every listing in that gallery is judged by Jev itself with the probabilities shown on each project page.

It is a community catalog on jevbooks.com, not published by TypeSafe AI, and it explains the patterns conceptually and links projects that adopted them. It reports no benchmarks or measured outcomes of its own, so the patterns are a vocabulary for reasoning about probabilities rather than measured results.

Questionnaire is flat fan-out, one thing per question and all of them at once in a single request. Prefetch is speculative heads: ask every branch up front and read only the one you take. Blind review is cross-check, where the reviewer never sees the answer being judged. Funnel is a cascade, cheap checks first and expensive ones last.