Jev Guides
Hands-on tutorials for the Jev judgment model: concepts, API calls, channels, confidence handling, cost and troubleshooting.
Jev API Channels: Official API vs OpenRouter (How to Choose)
Official Jev API vs OpenRouter compared: integration effort, billing, update speed and migration cost, plus a channel adapter and a migration runbook.
Updated 2026-09-21
Jev Cost and Latency: Real Pricing Math and Budgets
Jev cost and latency explained: real math at $0.0462 per 1M input tokens, LLM cost comparison, latency budgets and batching tips.
Updated 2026-09-21
Jev Troubleshooting: Common Errors and Proven Fixes
Jev troubleshooting guide: 401 and 404 model errors, 429 rate limits, JSON parse failures and confidence anomalies, with proven fixes.
Updated 2026-09-21
Jev API Tutorial: Make Your First Call in 10 Minutes
Jev API tutorial: step-by-step runbook from account to first call, copy-paste curl and Python examples, full request/response field walkthrough, failure mode table, cost measurement, and production checklist.
Updated 2026-09-24
How to Use Jev with Coding Agents (Cursor, Claude Code)
Use Jev API with coding agents like Cursor and Claude Code: skill file layout, complete worked example, retry rules, debugging a skill that won't trigger, and six anti-patterns.
Updated 2026-09-25
Jev Confidence Scores and Fallback Strategies That Work
Set Jev confidence thresholds that work: low-confidence fallback to LLMs, humans or retries, plus monitoring tips for production. Covers cold-start tuning, batch patterns, and anti-patterns to avoid.
Updated 2026-09-26
Jev AI vs LLMs: A Complete Comparison of Judgment vs Chat Models
Jev AI vs LLMs: capability boundaries, cost model, latency, determinism, explainability, regression testing, and a practical selection checklist for every pipeline step.
Updated 2026-09-27
Designing Jev Questions and State: A Practical Guide
Design stable Jev questions: option enumeration, boundary definitions, few-shot examples and how to pass context state between calls.
Updated 2026-09-28
Jev Model's Three Primitives: Judgment, Choice, Scoring
Jev's three primitives explained: judgment, choice and scoring questions, their request and response shapes, boundary conditions, cost trade-offs, and when to use each.
Updated 2026-09-29
How to Get a Jev API Key and Set Up Your Environment
How to get a Jev API key via OpenRouter, manage environment variables, set up billing and keep your key safe. A 10-minute setup guide.
Updated 2026-10-02
Jev vs LLM-as-a-Judge: Which Should You Use for Evaluation?
A practical comparison of jev ai vs LLM-as-a-Judge for automated evaluation pipelines: parsing pain, positional bias, per-grade cost arithmetic, and when a dedicated judgment model delivers more reliable scores at lower cost.
Updated 2026-10-06
Jev LLM Guardrails: Screen Every Input and Output in One Request
Build LLM guardrails with Jev: four hazard Noul questions plus a severity Score screened in a single request, then thresholded in your own code into pass, review, block or a support path. Covers input and output screening, two named policies and adapting the battery to your traffic.
Updated 2026-10-07
What Is Jev AI? A Working Guide to the Judgment Model API
What is Jev AI? A plain-English explainer of the Jev judgment model by TypeSafe AI — its three primitives, request shape, confidence field, cost profile, and when not to use it.
Updated 2026-10-07
Jev Language Support: English First, CJK Usable — How to Verify It on Your Own Content
Jev accepts natural-language text, but the official documentation is explicit that English is the primary training language and that other languages, CJK scripts included, are handled but not equally well. This guide turns that warning into a procedure: sample your own traffic per language, measure agreement and confidence separately, set per-language thresholds, and choose a fallback ladder before you ship.
Updated 2026-10-08
Jev Router: When Routing Beats Pinning Jev 1.13
Jev Router (typesafe/jev-router) picks a model and reasoning effort per request using Jev. Cost tiers, model preference lists, router metadata, error codes, and when to pin jev-1.13 instead.
Updated 2026-11-03
Jev API SDK Guide: Python and JavaScript Clients (typesafe-sdk)
Install and use the official TypeSafe SDKs to call the Jev API: typesafe-sdk and @typesafe-ai/sdk, client options, typed answers, RetryPolicy, zstd compression, logging redaction, model aliases and error codes.
Updated 2026-11-05
Jev API Access Routes: Official Endpoint, OpenCode Zen and OpenRouter
Three routes to the Jev API in November 2026: the official api.typesafe.ai endpoint, OpenCode Zen's jev-1.13 plus its free variant, and OpenRouter. Live endpoint probes, cost arithmetic per 1,000 requests, the no-credits key error, and a channel verification checklist.
Updated 2026-11-07
Jev AI Patterns: Speculative Fan-Out, Noul Screens and Companion Nouls
The six request-shaping patterns TypeSafe documents for the jev api: speculative fan-out, screening with a Noul first, companion Nouls, routing by confidence, routing by probability and composite scoring — with the worked examples, thresholds and recorded answer values from the official docs.
Updated 2026-11-09
Jev AI Structured Extraction: Pick Candidates, Read Dates, Verify with a Cascade
Extract structured data with the jev api without letting the model invent a value: regex candidates plus a Choice pick, seven-part date reading with the calendar math in code, and a verify-and-escalate cascade with Jev as verifier — using the recorded answer values and cost numbers from TypeSafe's extraction cookbooks.
Updated 2026-11-10
Jev Image Inputs and Box Questions: Sending Images and Reading Bounding Boxes
How to send images to the Jev API: the assets map and reference rules, the TypeSafe-Danger-Zone-Images header, where an image can go in state, instructions and criteria, Box questions and their four edges, measured OCR-check results, and the hard size limits.
Updated 2026-11-11
How to Use Jev in Cursor: Rules, Skills and a Thin MCP Wrapper
Wire the jev api into Cursor: .mdc project rules versus AGENTS.md, agent skills under .cursor/skills, and a thin MCP wrapper in .cursor/mcp.json — plus the confidence thresholds that belong in code, with the values and limits from TypeSafe's and Cursor's own docs.
Updated 2026-11-12
Jev AI Failure Modes: The 9 Jagged Edges TypeSafe Documents, and the Code That Fixes Each One
TypeSafe's own model-jaggedness page lists nine failure modes for the jev ai judgment model — literal reading, counting, date arithmetic, indirection, large state, adversarial content, contradictory criteria, choice option order and generation. This guide takes each one in turn with the documentation's stated cause, the fix it recommends, and a code-shaped workaround you can drop into a pipeline.
Updated 2026-11-13
Jev AI Function Calling: Turn Typed Signatures into Confidence-Scored Dispatcher Calls
TypeSafe's function calling cookbook maps a natural-language command onto one of ten typed Python functions by turning every closed-set argument into a Choice question and every optional argument into a Noul. This guide walks the measured output: 28 fillable arguments, 54 questions per command, and the confidence each of fourteen commands actually returned.
Updated 2026-11-14
Jev AI Hierarchical Classification: Beam Search Over Choice Probabilities
TypeSafe's hierarchical classification cookbook walks a taxonomy from root to leaf by asking one Choice question per node. This guide covers greedy versus beam search, the geometric-mean path score, the separation ratio, and the cookbook's measured result: beam search matched 3 of 3 expected leaves where greedy matched 2 of 3.
Updated 2026-11-17
Jev AI Infinite-Cardinality Choice: Picking One Item Out of 5,000 in Three Rounds
How to run a jev api Choice question when the option set is far bigger than the 255-option limit: a Noul filter over short descriptions, a re-read of the top eighty, and one Choice over twenty — with the official cookbook's measured 78% versus 33% for embedding search, and the 0.8 threshold that was right 48 times out of 48.
Updated 2026-11-18
Verify LLM Answers Against Their Source: Jev API Grounding Check (Three Nouls, One Request)
A grounding check for RAG answers built on the Jev API: three Noul questions in one request decide whether the retrieved paragraphs support what an LLM wrote, and turn into a single risk number. Measured on 1,000 MuSiQue answers with jev-1.14.0 on 2026-11-07: 73% passed the 0.2 threshold and 95.7% of those were right, against 88.5% if you keep every answer. All numbers, thresholds and demo scores are copied from the official verify_llm_outputs cookbook.
Updated 2026-11-19
Line-by-Line Semantic Search with the Jev API: One Choice Over 218 Line IDs Plus One Noul
Build document search that returns a quotable line instead of a paragraph: tag each line with an ID, let one Choice question score every ID, and gate the result with one Noul question in the same Jev API request. All numbers, thresholds and demo scores are copied from the official semantic_find cookbook, which searches GitHub's Terms of Service (218 lines, 43,980 characters) with jev-1.14.0.
Updated 2026-11-20