Jev Glossary

Every term in the judgment-only AI paradigm, in one line each — with a link to the page that treats it properly. Terms are ordered for reading top-to-bottom if you're new here.

01System One model

TypeSafe AI's category name for Jev, borrowed from Kahneman's Thinking, Fast and Slow: a model built for fast, intuitive, one-shot judgment (System 1) rather than slow deliberative reasoning (System 2). Input: structured state. Output: typed probabilistic decisions — never prose. Read more →

02RLCD

Reinforcement Learning for Calibrated Decisions — Jev's training method. Unlike RLHF/RLVR (which optimize for text humans or verifiers prefer), RLCD optimizes for calibration: a stated 0.8 probability should be correct ~80% of the time. Verify it on your own data before trusting it. Read more →

03Noul

The yes/no primitive. A Noul question returns a single probability in [0, 1] and — unlike Choice and Score — no separate confidence field. Use it for binary decisions: spam?, phishing?, needs a human? Read more →

04Choice

The pick-one-of-N primitive. You supply up to 255 options; Jev returns the selected option, a per-option probabilities distribution, and a confidence value. Use it for routing, tagging, and intent buckets. Read more →

05Score

The rating primitive. You define a scale of 2–10 levels; Jev returns a score, a legend describing the levels, per-level probabilities, and a confidence value. Use it for grading and ranking. Read more →

06state

The structured record Jev judges — a string, object, or array sent with every request. Not a prompt: think of it as the case file you hand an expert. Filter fields in code first; irrelevant detail measurably degrades accuracy. Read more →

07questions

The typed map of judgments requested in one API call. Each value is a Noul, Choice, or Score question with required instructions and optional criteria. The map keys are your own names and are never sent to the model — and yes, you can mix primitives in one request. Read more →

08jaggedness

The unevenness of Jev's accuracy: strong on some tasks and framings, weaker on others, and not always self-consistent across primitives. The practical consequence: calibrate on your own data, and treat reframing instability as a "route to fallback" signal. Read more →

09calibration

The property that stated probabilities match reality: among all 0.8 answers, ~80% should be correct. Calibration is what makes threshold-based automation ("auto-approve above 0.95") meaningful — and why it must be measured, not assumed. Read more →

10speculative fan-out

The pattern of packing many small questions about the same item into one request: one state transmission, one HTTP call, many typed verdicts back. The cheapest latency and cost optimization in the Jev toolkit. Read more →

11Jevons paradox

William Stanley Jevons' 1865 observation that making a resource dramatically more efficient increases total consumption of it. Jev is named for the bet that radically cheaper machine judgment won't reduce judging — it will put judgment everywhere it was previously too expensive to apply. Read more →

Definitions follow community documentation (learnjev.com, jevai.wiki, jev101.com) cross-checked against public early-access coverage. Unofficial fan-made handbook — not affiliated with TypeSafe AI or jev.com.