Jev glossary

Every term the course uses, in the sense the course uses it. Alphabetical.

A to F

Act / confirm / human: The three outlets every automated decision should have. Act runs unattended, confirm asks a person or the customer before a risky action, human hands the case over. Lesson 4.
Alias: A model name that points at a version and moves over time. jev-latest and jev-preview both resolve to jev-1.13.0 as of 2026-09-19. Pin the version in production.
Benchmark (community): The Flow Dynamics AI brief of 2026-09-17: 40 real German call transcripts, 39 labelled correctly by Jev, one confident miss (call 028). Published at deep-cosmos-bhsx.here.now.
Call 028: The benchmark's blind spot: an incomplete transcript labelled "not sales" at 0.93 where the right answer was "unclear", an option the question set did not offer. The course's standing example for escape options and confidence gating.
Choice: Primitive that picks one option out of up to 255 and returns a probability per option plus a confidence. For "which kind of situation is this".
Composite scoring: Pattern: several concrete answers combined in code with explicit weights, instead of asking the model for a single rating. Lesson 5.
Confidence: A 0..1 statistic on Choice and Score answers describing how peaked the probability distribution is. Falls faster than the top probability as mass spreads. Not a guarantee. Lesson 4.
Context rot: Accuracy loss when the relevant sentence sits inside a large amount of irrelevant state. Fix by trimming.
Criteria: The options of a Choice (a map of name to description or null) or the levels of a Score (an ordered list). May be structured objects with what, not_for and examples.
Decision model: A model that returns calibrated probabilities over options you define, instead of generated text. Jev is one. Contrast: language model.
Deep link (Playground): A URL that opens the TypeSafe Playground with state and questions pre-filled, so an exercise can be run on the learner's own account without an API key.
Escape option: An option such as unclear, other or none that gives the honest answer somewhere to go. Missing in the benchmark's question set; see call 028.
Fan-out (speculative): Pattern: ask every question the workflow might need in one request and let code choose. Cheap because questions run in parallel and output is free. Lesson 5.
Floor: The confidence below which no automated action fires and the case goes to a human. The docs' examples start at about 0.5; your shadow run sets the real value.

G to P

Guardrail: A Choice or set of Nouls placed in front of an LLM agent to decide scope, safety or language before the agent runs.
High bar: The confidence a risky (hard to undo) action needs before it runs unattended. The docs' examples use 0.85 to 0.9.
Instructions: The question text of a primitive. Must agree with the criteria; contradictions make answers unpredictable.
Intent routing: Pattern: a Choice decides what kind of request this is, and each kind has its own handler (code, a narrow LLM, a human). Lesson 5.
Invariant (structural): A relationship between answers that a logician would expect, such as noul(A) + noul(not A) = 1. Jev does not guarantee any. Ask each fact once.
Jaggedness: TypeSafe's term for the uneven ability profile of a model version, published per version. For jev-1.13: counting, dates, literal reading, indirection, large state, adversarial content, contradictions, invariants, generation.
Jev: TypeSafe AI's decision model family. Current version jev-1.13.0. Served through the System One endpoint.
Legend: In a Score answer, the list of level descriptions echoed back so the score index can be read without the request.
Level: One rung of a Score ladder, indexed from 0. Described as a situation, never as a number.
Noul: Primitive that returns the probability, 0..1, that one statement about the state holds. No confidence field. Phrase positively, one fact per Noul.
Pin (a version): Specifying "model": "jev-1.13.0" instead of an alias so thresholds keep their meaning across model releases.
Playground: The web console at console.typesafe.ai/decode where a request can be typed, run and its raw response read. The course's IDE.
Primitive: One of the three question types: Choice, Score, Noul.
Probabilities: In a Choice answer, the distribution over options (sums to 1). In a Score answer, the distribution over levels. Use them, not only the winner, in composites.

Q to Z

RLCD: Reinforcement learning from calibrated decisions, the training approach TypeSafe describes for Jev: rewarded for calibrated probabilities on decisions rather than for fluent text.
Risky action: An action that is expensive or impossible to undo: refund, booking, blocking, closing. Gets the high bar instead of the floor.
Score: Primitive that places the state on an ordered ladder of 2 to 32 levels and returns the expected level index as a float, plus the distribution and a confidence.
Shadow evaluation: Running the pinned model on live traffic for a period without acting on it, logging answers next to what humans did, then tabulating accuracy per confidence band to set thresholds. Lesson 6.
State: The input to a request: a string, or a JSON object or array. Up to 32k tokens. Referenced from instructions with backtick paths.
System One: The API endpoint name (/v1/systemone) and TypeSafe's name for the product: fast, non-generative judgement, after Kahneman's term.
Threshold: A confidence value at which routing changes outlet. Two per decision: floor and high bar. Kept in one versioned file with the questions.
Usage: The token counts returned with every response. Only input tokens are billed, at $0.042 per million.
Wikirace: TypeSafe's demo video in which a Jev-driven agent races LLM agents across Wikipedia, choosing one link per hop from a Choice over all links on the page. Lesson 5.