Jev glossary
Every term you meet while building with Jev, defined in plain language and grounded in TypeSafe's documentation: the three primitives, confidence and calibration, model IDs and pricing, the named patterns, and the jagged edges the docs admit to.
Each entry links to the resources and docs behind it. Definitions track the current jev-1.13.0 documentation, last reviewed .
- Terms
- 117
- Categories
- 7
- Linked resources
- 62
- Docs links
- 117
Browse by category
7 groups, from core concepts to the ecosystem around the API.
- Core Concepts20The vocabulary behind System One: what Jev is, what state and questions are, and the architecture TypeSafe recommends.
- Primitives15Choice, Score, and Noul, plus the instructions, criteria, and answer fields you use to define them.
- Answers & Confidence10How probabilities, confidence, calibration, and thresholds turn typed answers into decisions code can gate.
- API & Models21The endpoint, model IDs and aliases, context limits, token pricing, retries, and error handling.
- Patterns23The named compositions the docs and cookbooks use: fan-out, confidence gates, cascades, re-ranking, and more.
- Evaluation & Training21RLCD and its alternatives, the documented jagged edges, and how Jev's quality is measured.
- Ecosystem7The SDKs, skills, MCP servers, gateways, and community ports that put Jev inside your stack.
All 117 terms
Search the glossary, filter by category, or jump to a letter.
117 terms
A
- Adversarial contentPrompt injectionJev treats state as data, not as hostile input; injected instructions can move answers, so write precise criteria and test edge cases.Evaluation & Training
- Agent loopLLM agentAgentic loopA control pattern where the model chooses its next step repeatedly; TypeSafe's guidance is to prefer a code-owned workflow instead.Core Concepts
- Agent skillTypeSafe agent skillA drop-in SKILL.md that teaches coding agents to structure Jev states and questions and to batch questions per call.Ecosystem
- AI gatewayA routing layer that serves Jev alongside other models; Vercel AI Gateway lists typesafe-ai/jev as a hosted model.Ecosystem
- AI primitivePrimitiveOne of TypeSafe's small, typed building blocks — Choice, Score, Noul — that pair a question with a typed answer you compose in code.Core Concepts
- AI-powered softwareThe architecture TypeSafe recommends: code owns the workflow and the model appears only where programmable common sense is needed.Core Concepts
- AnswerTyped answerThe structured result returned under a question's ID: a choice, a score, or a probability of yes, plus distributions and confidence.Core Concepts
- API errorThe four documented failure statuses — 401, 422, 429, and 529 — with retry guidance for the last two.API & Models
- API keyThe bearer credential for the TypeSafe API, created in the console and read from TYPESAFE_API_KEY by the official SDKs.API & Models
- Atomic questionA question that asks one specific, well-scoped thing — the kind of judgment a knowledgeable person makes in a few seconds.Core Concepts
- AutoresearchAutoresearch feature discoveryA loop that proposes questions, turns free text into numeric features, and uses model errors to improve a classical regressor.Patterns
B
- Backtick path referenceDot-and-index pathNaming a state field inside instructions with backticked dot-and-index paths, such as `ticket.messages[0].text`.Primitives
- Benchmark claimsTypeSafe's headline multipliers — 193.6x faster, 444.6x cheaper, a 238x lower input price — come from its own workflow comparisons.Evaluation & Training
C
- CalibrationThe property that outcomes given a probability of 0.2 occur about 20% of the time, measured across many predictions.Answers & Confidence
- ChoiceChoice questionPick one option from a set you define; returns the selected option, a probability for every option, and confidence.Primitives
- Choice optionOne key in a Choice criteria map, with a description that can be a string, a structured object, or null.Primitives
- Client SDKTypeSafe SDKThe official typed clients for the API — Python and JavaScript/TypeScript — with retries and inferred answer types.API & Models
- Coding agentThe fastest-growing corner of the Jev ecosystem: MCP servers, tool-call gates, context pruners, and skill rankers.Ecosystem
- Community portAn unofficial client or integration for a language or framework TypeSafe does not ship officially, from Go and .NET to DSPy.Ecosystem
- Composable AITypeSafe's manifesto position: intelligence as a dependable primitive that software branches on, rather than an assistant that keeps humans in the loop.Core Concepts
- Composite scoringBreak a judgment into atomic Score questions, normalize each, and combine them with weights you control in code.Patterns
- ConfidenceA 0-1 number on every Choice and Score answer that collapses the probability distribution into one certainty score.Answers & Confidence
- Confidence floorA baseline below which code never acts on an answer, regardless of what the answer says.Answers & Confidence
- Confidence gateA threshold on confidence that decides whether code acts, confirms first, or escalates to a person or a bigger model.Answers & Confidence
- Confidence-gated routingUse the answer to decide what to do and confidence to decide whether to act, with per-action thresholds set by the cost of being wrong.Patterns
- Consensus labelsReference answers built from two frontier models at high thinking, used as the ground truth in TypeSafe's workflow evals.Evaluation & Training
- Context rotThe accuracy loss that comes from padding state with detail unrelated to the question; filter state before sending it.Core Concepts
- Context windowThe 64k-token budget covering state plus all questions, with a 32k sub-budget for state plus the longest question.API & Models
- Contradictory criteriaWhen instructions and criteria ask for different things, accuracy drops; treat criteria as an extension of the instruction.Evaluation & Training
- CookbookAn end-to-end recipe in the TypeSafe docs that shows a real problem, with its dataset, latencies, and measured numbers.Ecosystem
- Cost/quality frontierPlotting accuracy against cost and time on log scales, where up and to the left is better; cascades are designed to sit there.Evaluation & Training
- CriteriaThe possible answers for a question: a map of options for Choice, an ordered list of levels for Score, or true/false for Noul.Primitives
D
- Dates as textJev reads dates as strings, not ordered quantities; extract components with Choice and do the ordering and math in code.Evaluation & Training
- DecompositionBreaking a broad judgment into one question per factor, then combining the answers with weights and rules you control in code.Core Concepts
E
- Entity alignmentEntity matchingsameAsDeciding whether two records describe the same entity, with one Score per pair plus companion Nouls for the fields that disagree.Patterns
- Entry typeThe value shapes accepted throughout a request — string, object, array, or null — for state, instructions, and criteria fields.Primitives
- Exponential backoffThe retry strategy the docs recommend for 429 and 529 responses: wait longer after each attempt instead of retrying immediately.API & Models
F
G
- Generation limitationJev is not trained to generate text; forcing it by chaining choices is slow and unreliable, so extraction becomes a Choice instead.Evaluation & Training
- GuardrailScreening messages going into or out of an LLM app with hazard Nouls and severity Scores, then thresholding the probabilities.Patterns
H
- Hierarchical classificationWalk a deep taxonomy level by level, using each Choice answer to pick the next level's options, keeping several paths alive.Patterns
- Human reviewEscalationThe explicit path for uncertain or high-stakes cases: route to a person instead of acting on a low-confidence answer.Patterns
I
- IndirectionQuestions about a property of a property, or with double negatives, cost accuracy; write instructions as directly as possible.Evaluation & Training
- Input tokensThe tokens billed on every request: state plus questions, including speculative questions your code ignores.API & Models
- InstructionsThe field that states what a question should judge; accepts a string, an object, or an array for extra structure.Primitives
- Intent routingClassify intent with a Choice and complexity with a Score, then send each branch to deterministic code, a specialist model, or a person.Patterns
J
- JaggednessThe documented set of jev-1.13 failure modes — literal reading, counting, dates, indirection, context rot — each with a guardrail.Evaluation & Training
- Jevjev-latest modelTypeSafe's flagship modelTypeSafe AI's flagship model and the first System One model: typed decisions and calibrated probabilities instead of generated text.Core Concepts
- jev-1.13.0Jev 1.13The current versioned Jev build: $42 per billion input tokens, a 64k context, and 250,000 tokens per second.API & Models
- jev-latestThe default model alias in the SDKs and docs: the most recent stable official Jev release, currently jev-1.13.0.API & Models
- jev-previewThe alias for the most recent Jev release whether or not it is official; currently identical to jev-latest.API & Models
L
- Language supportEnglish is Jev's primary training language; other languages are accepted but less accurate, so test before relying on them.API & Models
- Line-by-line searchSemantic findRank every line of a document in one request with a Choice, while a Noul checks whether the document contains an answer at all.Patterns
- Literal readingJev answers the question you wrote, not the one you meant; scoping words, negations, and implied conditions are read at face value.Evaluation & Training
M
- Machine Native IntelligenceTypeSafe's name for AI with software-like properties: structure, reliability, observability, testability, speed, consistency, and low cost.Core Concepts
- Machine-to-machine automationM2MAI-to-AIAutomation where software calls AI directly, expected to be roughly 99% of large-scale AI interactions versus 1% human-facing.Core Concepts
- MCP serverModel Context Protocol serverA server that exposes Jev to coding agents as a tool through the Model Context Protocol, usually one evaluate call.Ecosystem
- Mode droppingPreference optimization narrowing a model's output distribution toward a favored style and away from other valid outputs.Evaluation & Training
- Model aliasA name like jev-latest that resolves to a versioned model ID and can move when a new release ships.API & Models
- model fieldThe required request field that selects which model handles the call; accepts versioned IDs and aliases.API & Models
N
- Neuro-symbolic AINeuro-symbolicThe manifesto's framing of TypeSafe's aim: neural networks for perception paired with symbolic logic, or smart if-statements.Core Concepts
- NormalizationScaling each Score by its number of levels — dividing by len(criteria) - 1 — before weighting it in a composite.Patterns
- NoulNoul questionYes/no questionThe yes/no primitive: one probability from 0 to 1, where the value is the answer and no separate confidence is returned.Primitives
- Noul criteriaOptional true and false descriptions that define what a yes and a no mean for a Noul question.Primitives
- Noul probabilityP(yes)The single 0-1 value a Noul returns: the probability that the answer is yes, with no separate confidence field.Answers & Confidence
- Null criteriaPassing null as a Choice option description when the option name needs no extra explanation.Primitives
O
P
- Parallel beam searchBeam searchKeep the top-k branches at each taxonomy level instead of only the winner, so a wrong first choice can still be recovered.Patterns
- Parallel questionsBatching many questions into one request instead of many calls, which is dramatically cheaper and faster with identical answers.Patterns
- PlaygroundThe browser workbench for pasting state and typed questions and inspecting answers, probabilities, confidence, latency, and usage.API & Models
- Pre-parsed value extractionRegex finds candidate spans, then Jev picks the requested one so code copies a verbatim value that cannot be invented.Patterns
- Primitive limitsThe hard caps to design around: 255 Choice options, 2 to 10 Score levels, and a 64k context covering state plus questions.Primitives
- Probability distributionprobabilitiesThe floats summing to 1 across your options or levels that every Choice and Score answer returns under probabilities.Answers & Confidence
Q
R
- RAG passage classificationScore each retrieved passage with four Noul questions, then decide in code which passages reach the answering model.Patterns
- Rate limitThe throughput caps — 250,000 tokens per second and 1,200 requests per minute — that return 429 when exceeded.API & Models
- Re-rankingScore each query-candidate pair with one question and reorder a shortlist; the CLERC example lifts top-1 accuracy from 5% to 18%.Patterns
- RepeatabilityHow stable answers are across repeated evaluations, measured separately from accuracy and reported as probability standard deviation.Evaluation & Training
- Retry policyThe SDK configuration for retrying failed requests: attempt count, retryable statuses, backoff, and retry-header handling.API & Models
- RLCDReinforcement learning for calibrated decisionsTypeSafe's post-training method: reinforcement learning that optimizes for calibrated decisions and probabilities instead of generated text.Evaluation & Training
- RLHFReinforcement learning from human feedbackThe human-preference training method that turned pretrained models into chatbots, and the method TypeSafe deliberately did not use.Evaluation & Training
- RLVRReinforcement learning with verifiable rewardsTraining against automatically verifiable rewards, which produced strong reasoning models that are slower and more expensive.Evaluation & Training
S
- ScoreScore questionRate the state against 2 to 10 ordered levels; returns a probability-weighted position, per-level probabilities, and confidence.Primitives
- Score legendThe Score answer field that maps each level number back to its description, so results stay readable after the fact.Primitives
- Score levelOne entry in a Score's ordered criteria list; levels are numbered by array position from 0 and judged independently.Primitives
- SDE cascadeStructured-data-extraction cascadeA two-stage extraction cascade: a mini model extracts, a Noul battery verifies, and a confidence gate escalates to a reasoning model.Patterns
- Second requestA follow-up call that is only justified when code cannot build the second request until it has the first answer.Patterns
- Self-consistencyRepeatability across repeated evaluations: Jev's probability standard deviation measured at about 0.01 in the consistency cookbooks.Patterns
- Skill suggestionSkill rankingRank a large skill catalog in one request, then re-read the top few in full detail and pick at most one.Patterns
- Smart if statementTypeSafe's shorthand for what a System One decision is: a condition your code can branch on directly, backed by a model judgment.Core Concepts
- Speculative fan-outAsk every question the system might need in one request — including ones only some branches use — and let code discard the rest.Patterns
- Speculative questionA question whose answer only matters on some code paths, asked anyway because questions are cheap and parallel.Patterns
- StateThe content a request asks Jev to evaluate: a string, a named object, or an array of messages, shared by every question in the request.Core Concepts
- Structural invariantsArithmetic identities you might assume — complementary probabilities summing to 1, agreement across question types — are not guaranteed.Evaluation & Training
- Structure recoveryAutoformatReconstruct Markdown from unformatted text in two requests: stitch split lines, then classify every block's type.Patterns
- Structured instructionsStructured criteriaPutting the question in one field and its supporting data, examples, or focus notes in others instead of one long string.Primitives
- SycophancyThe tendency of preference-trained models to say what people want to hear, one of the reasons TypeSafe chose a calibration objective.Evaluation & Training
- System OneSystem One modelThe class of AI models built to make fast, structured decisions software can use directly, named after Kahneman's fast-thinking System 1.Core Concepts
- System One adapterThe official drop-in TypeSafeClient replacement backed by OpenAI, Anthropic, and compatible LLM APIs.Ecosystem
- System One endpointPOST /v1/systemoneThe single HTTP endpoint that evaluates a state against typed questions: POST https://api.typesafe.ai/v1/systemone.API & Models
- System Two taskWork that needs extended, multi-step reasoning; the docs treat it as a signal to decompose rather than a job for Jev.Core Concepts
T
- Text-only inputJev evaluates strings, JSON objects, and arrays of text; images, audio, and video are not supported and must be pre-processed.Core Concepts
- The Bitterest LessonTypeSafe's essay extending Sutton's lesson: doing the right task beats data, which beats compute, which beats algorithms.Evaluation & Training
- ThresholdThe value your code compares against a score, a Noul probability, or a confidence to turn an answer into an action.Answers & Confidence
- Token pricingJev's published rate: $42 per billion input tokens, with output tokens free.API & Models
- Typed decisionA model output that conforms to a software type your code expects, rather than free-form text that has to be parsed.Core Concepts
- TypeSafe AITypeSafeThe company behind Jev and the System One API, including the official SDKs, cookbooks, console, and published workflow evals.Core Concepts
- TypeSafe ConsoleThe web console for API keys, model management, and the playground, with walkthrough lessons and realistic use cases.API & Models
U
- UncertaintyThe honest "I'm not sure" signal that lets a system defer instead of guessing, exposed through probabilities and confidence.Answers & Confidence
- Uncertainty bandReview bandA middle range of probabilities, such as 0.30 to 0.70, that maps to an explicit uncertain outcome instead of a hard yes or no.Answers & Confidence
- Unreliable countingCountingJev does not count reliably — characters, occurrences, or list items — and error grows with the size of the thing counted.Evaluation & Training
- UsageThe response field reporting input and output token counts for a request, used for cost tracking.API & Models
W
- Weighted scoreScore expectationThe Score answer computed as each level number times its probability, which is why a score can land between two levels.Answers & Confidence
- Workflow evalsTypeSafe's published evaluations of four automation workflows, comparing Jev and frontier LLMs as structured workflows versus single prompts.Evaluation & Training