$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people
~/wiki/syntheses

Glossary

[ synthesis ][ updated 2026-09-17 ][ confidence high ][ jev-1.13.0 ][ python sdk 0.6.0 ][ js sdk 0.6.0 ]#glossary · terminology · definitions · reference · index

TL;DR One alphabetical table. Terms that are literal API or SDK identifiers are in backticks and spelled exactly as the wire format spells them; everything else is a concept, pattern, or company term. Follow the link for the contract, the numbers, and the gotchas.

Term Definition See
401 / 422 / 429 / 529 The four documented HTTP failures: bad or missing key, request-body validation error, rate limit exceeded, and service overloaded. 429 and 529 are retryable with backoff; 401 and 422 are not. HTTP status codes, rate limits, retry semantics
Agent skill (typesafe-ai) TypeSafe's skill for coding agents, shipped as the Claude Code plugin typesafe (version 0.5.7, MIT) and via npx skills add. Its central instruction is that the live docs, not the skill, are the source of truth. The typesafe-ai agent skill and Claude Code plugin
Alias A moving model name — jev-latest (most recent stable release, the SDK default) and jev-preview (most recent release of any kind). Both currently resolve to jev-1.13.0; pin the versioned ID if you have tuned thresholds. Models, aliases, pricing, rate limits, context
Answer The typed value returned for one question, carrying the same type as the question that produced it. Choice and Score answers add probabilities and confidence; a Noul answer is a single number. Primitives: Choice, Score, Noul
answers The response map keyed by the question ids you chose. It replaced preview's ordered responses array, so v1 code looks answers up by key, never by position. HTTP API: POST /v1/systemone and GET /v1/models
Btok / Mtok TypeSafe's billing units: a Btok is a billion tokens, an Mtok a million. Jev is $42 per Btok / $0.042 per Mtok on input tokens, with output tokens free. Models, aliases, pricing, rate limits, context
Calibration The property that across many predictions, outcomes assigned probability 0.8 occur about 80% of the time. It is a statement about groups of answers and guarantees nothing about any single one. AI primer: why calibrated decision models
Cardinality ceiling The launch post states Jev "supports a cardinality up to 255"; above that, TypeSafe used a two-stage system of independent scoring followed by an explicit choice, "hence the occassional slowdown". Choice questions
Choice The primitive that selects one option from a set you define in criteria. It returns the winning option, a probability distribution over every option, and a confidence. Choice questions
choice The Choice answer field holding the highest-probability option name. In the preview API this field was called chosen. HTTP API: POST /v1/systemone and GET /v1/models
Composite scoring The pattern of asking several independent Score questions in one call, normalising each to 0–1, and combining them with weights your code owns — so changing priorities means changing a coefficient, not a prompt. Composite scoring
Confidence A 0–1 statistic computed from an answer's probabilities that summarises how concentrated the distribution is. Present on Choice and Score answers only; the formula is not published, so compute your own from probabilities if you need a specific measure. Confidence vs probability
Confidence-gated routing The pattern of treating confidence as a second decision axis: a floor below which nothing is automated, then a per-action threshold sized to the consequences of being wrong. Confidence-gated routing
Consensus labels TypeSafe's term for the reference answers in its workflow evals, produced by averaging GPT-6 Astra and Claude Fable 5.1 at high thinking. "Accuracy" in those evals means agreement with this average, not correctness. Workflow evals: how TypeSafe measures Jev
Console https://console.typesafe.ai — login, the Playground, and API keys at /settings/keys. The MCA calls it the "Web Interface". console.typesafe.ai (console + playground)
Context budget 64k tokens per request for state plus all questions, and 32k tokens for state plus the single longest question. Because state counts in both, a large state shrinks both headrooms. Models, aliases, pricing, rate limits, context
Cookbook One of eighteen worked upstream recipes: a state, a set of questions with their exact instructions and criteria, and the composition logic in ordinary Python. Cookbooks overview
cooksafe The helper package every published cookbook installs from https://pypi.typesafe.ai/, which returned 404 publicly as of 2026-09-17. Nothing in the recipes depends on it — install typesafe-sdk and reimplement JsonCache and make_playground_link. Cookbooks overview
Credits The metered balance the MCA bills usage against; purchased Credits expire at the earlier of end of Term or 12 months after purchase, and the consumption rate "may vary based on account settings, including the model used". Legal: MCA, DPA, privacy, data retention
criteria The field that defines a question's answer space: a map of option → description for Choice, an ordered array of level descriptions for Score, and an optional {true, false} clarification for Noul. Primitives: Choice, Score, Noul
DPA TypeSafe's Data Processing Addendum (last updated 2026-04-24), incorporated into the MCA and controlling where the two conflict. It sets the processor role, 72-hour breach notice, annual audit right, and the EU/UK/Swiss transfer mechanisms. Legal: MCA, DPA, privacy, data retention
Early access Jev's availability status since 2026-09-15: waitlisted, not open self-serve signup. Jev (model)
EntryType The JSON type every prompt-bearing field accepts — string, object, array, or null — so an instruction or an option description can be a structured object rather than a sentence. Structured instructions, options, levels, criteria
Expected score What a Score answer's score actually is: the probability-weighted mean of the level numbers, which is why it can fall between levels and why two different distributions can produce the same value. Do not interpolate it into a real-world magnitude. Score questions
Fan-out (speculative) The pattern of sending every question your decision tree might need in one call — including branches you have not taken — and discarding the answers code does not use. Questions run in parallel, so extra ones cost tokens rather than latency. Speculative fan-out
instructions The field carrying the judgment you want made. Marked required by the API docs for all three question types, though the OpenAPI schema lists it as nullable — always send it. HTTP API: POST /v1/systemone and GET /v1/models
Intent routing The pattern of classifying a request with one cheap Jev call, then dispatching to the cheapest adequate handler: plain code, a specialist LLM, or a human. Intent routing
Jaggedness TypeSafe's term for a model's uneven capability surface. The jev-1.13 jaggedness page documents nine failure modes and the mitigation for each. Jev 1.13 jaggedness: known failure modes
Jev TypeSafe's first and flagship System One model, named after William Stanley Jevons. Text in, typed decisions with calibrated probabilities out. Jev (model)
jev-1.13.0 The current versioned model ID as of 2026-09-17, and what both aliases resolve to. The response's model field reports the version that actually answered. Models, aliases, pricing, rate limits, context
JsonCache The cookbook helper that memoizes results into a JSON file keyed on the call's arguments — including a rubric hash and the model name — so republished cookbook numbers reproduce without API spend. Cookbooks overview
legend The Score answer field mapping each level number (as a string key) back to the criteria description you supplied, so you can label a score without keeping the rubric around. New in v1. Score questions
Literal reading Jaggedness failure mode 1: jev-1.13 answers the question you wrote rather than the one you meant, taking scoping words, negations, and implied conditions at face value. Jev 1.13 jaggedness: known failure modes
Machine-Native Intelligence TypeSafe's term for AI with software-like properties — structure, reliability, observability, testability, speed, consistency, low cost — as opposed to chat-shaped AI aimed at a human reader. AI primer: why calibrated decision models
MCA The Master Customer Agreement (last updated 2026-08-27), TypeSafe's commercial contract. It defines Input, Output, Telemetry, the acceptable-use restrictions, the liability cap, and the absence of any uptime SLA. Legal: MCA, DPA, privacy, data retention
Mode dropping The RLHF side effect TypeSafe cites as motivation for RLCD: preference optimisation narrows a model toward a favoured style, a milder version of GAN-style mode collapse. AI primer: why calibrated decision models
model The required request field naming the model or alias to use, normally "jev-latest". The same field on the response reports which version answered, and the two can differ. HTTP API: POST /v1/systemone and GET /v1/models
Noul The yes/no primitive. It returns a single probability that the statement is true and carries no confidence field; criteria.true / criteria.false optionally clarify what yes and no mean. Noul (yes/no) questions
noul The Noul answer field: a number from 0 (no) to 1 (yes), with values near 0.5 meaning genuinely uncertain rather than "medium". Called probability in the preview API. Noul (yes/no) questions
nouls / choices / scores The three cached typed views on a Python SystemOneResponse, alongside the documented answers map. Upstream samples use both spellings. Python SDK responses, answers, usage, models
OpenAPI document The machine-readable contract at https://docs.typesafe.ai/openapi.json: openapi: 3.1.0, info.title: "TypeSafe", info.version: 0.2.0, two paths and sixteen component schemas. OpenAPI component schemas
Parallel sampler The architectural claim behind Jev's speed: all outputs are produced in a single query rather than one token at a time conditioned on the last, which is why adding questions barely changes latency. System One Models
Playground The console's interactive surface at /playground, where a state and question set can be tried and shared as a #share/... or ?share=shr_... link — the form the cookbooks' share links take. console.typesafe.ai (console + playground)
Preview endpoint The retired POST /preview/evaluation API, replaced by POST /v1/systemone. It used document, a prompts array, options/levels, and a responses array, and its confidence used a different computation. Migrating from /preview/evaluation to /v1/systemone
Primitive One of the three question/answer pairs — Choice, Score, Noul — that make up every Jev request. You compose their typed answers in code rather than asking one big question. Primitives: Choice, Score, Noul
probabilities The distribution an answer returns over your Choice options or Score levels, summing to approximately 1. It is the raw signal; confidence is one summary of it, and you can compute others. Confidence vs probability
Question One judgment for the model to make about the state, defined by a type, instructions, and usually criteria. Every question in a request sees the same state and is evaluated independently. Primitives: Choice, Score, Noul
Question ID The key you choose in the questions map, which the answer comes back under. It is never sent to the model, so the full meaning must live in instructions. HTTP API: POST /v1/systemone and GET /v1/models
retry-after / retry-after-ms Response headers naming how long to wait before retrying. Both SDKs prefer retry-after-ms when present and cap any honoured server delay at 60 seconds before falling back to their own backoff. HTTP status codes, rate limits, retry semantics
RetryPolicy The SDK object controlling retries: by default 2 retries, 500 ms initial backoff doubling to a 5,000 ms cap, 25% jitter, retryable statuses {408, 429, 500–599}. Since 0.6.0 invalid values raise at construction. Python SDK retries, exceptions, constants
RLCD Reinforcement Learning for Calibrated Decisions — TypeSafe's post-training method, presented as a third path alongside RLHF and RLVR, optimising probabilities against outcomes instead of text against preference. AI primer: why calibrated decision models
RLHF Reinforcement learning from human feedback — the method that turned pretrained models into chatbots, co-invented by TypeSafe cofounder Diogo Almeida, and the one TypeSafe positions itself against. AI primer: why calibrated decision models
RLVR Reinforcement learning with verifiable rewards — the method behind reasoning models strong at programmatically checkable tasks, at higher latency and cost. AI primer: why calibrated decision models
Score The primitive that rates a state against an ordered array of described levels, returning a probability-weighted position plus legend, probabilities, and confidence. Score questions
score The Score answer field: a float between 0 and len(criteria) - 1. Called expectation in the preview API. HTTP API: POST /v1/systemone and GET /v1/models
state The single input every question in a request is evaluated against — a string, JSON object, or array of text values, and nothing else. Unrelated material in it costs accuracy. State: what you send Jev
System 1 / System 2 Kahneman's distinction in Thinking, Fast and Slow between fast intuitive judgment and slow deliberate reasoning, from which TypeSafe takes the name "System One". System One Models
System One LLM wrapper TypeSafe's harness that constrains an LLM to emit System One-compatible structured decisions, used for the LLM baselines in its workflow evals. TypeSafe notes it "tends to be slower and more expensive" than unconstrained decisions. Workflow evals: how TypeSafe measures Jev
System One model The model class: evaluates one state, returns typed answers and calibrated probabilities, never generated text, with the answer space fixed in advance. System One Models
system-one-adapter The open-source Python package (v0.1.4) that answers typesafe_sdk questions with an OpenAI or Anthropic model instead of Jev, returning the same answer objects plus latency, token and per-attempt traces. system-one-adapter: LLM-backed drop-in for TypeSafeClient
Telemetry The MCA's carve-out category — "technical logs, hashes, summary statistics and classifications, metrics, and learnings" — which TypeSafe may process without restriction and which survives termination. Legal: MCA, DPA, privacy, data retention
TYPESAFE_* env vars Exactly four are read by the SDKs: TYPESAFE_API_KEY, TYPESAFE_BASE_URL, TYPESAFE_DEFAULT_MODEL, TYPESAFE_LOG_LEVEL. TYPESAFE_MODEL, TYPESAFE_PRICE and similar names appear only in cookbook code. TYPESAFE_* environment variables across SDKs
typesafe-client The legacy Python package for the preview API. Every release (0.1.x and 1.0.x) sends document and no longer works against the API — replaced by typesafe-sdk. Migrating from /preview/evaluation to /v1/systemone
typesafe-sdk / @typesafe-ai/sdk The official Python (PyPI, import typesafe_sdk) and JavaScript/TypeScript (npm) clients, both at 0.6.0 and both MIT. Note the unrelated PyPI package typesafe is not TypeSafe AI's. Python SDK: install, clients, system_one()
Uncertain band The application-side pattern of mapping a middle range of probability or confidence to an explicit third outcome — human review — instead of forcing yes/no. TypeSafe's cookbooks use 0.30–0.70 for Nouls and a 0.60 top-probability floor for Choices. Cookbook: Self-consistency — nouls
usage The response object reporting input_tokens and output_tokens. Only input tokens are billed; it replaced preview's placeholder billing_units. HTTP API: POST /v1/systemone and GET /v1/models
v1 API The current stable API: POST /v1/systemone and GET /v1/models under https://api.typesafe.ai, with state, a questions map, and an answers map. HTTP API: POST /v1/systemone and GET /v1/models
Workflow eval TypeSafe's published evaluation format: a fixed code workflow, every model run through the identical harness, scored against consensus labels. Source of the "193.6x faster, 444.6x cheaper" claim. Workflow evals: how TypeSafe measures Jev
ZDR Zero data retention, offered "for enterprise customers" via privacy@typesafe.ai. Outside ZDR no retention period is stated anywhere in the legal documents. Legal: MCA, DPA, privacy, data retention

Related

Sources