Where to start

The list below is one path through the digest. Every figure on those pages is a claim from the people who published it unless a third party ran the same test.

1. What it returns

What Jev returns is the explainer: Choice, Score, Noul, and the claim that the model does not write text. The launch story is Almeida's September 15 announcement and TypeSafe's own workflow table. Jaggedness is the company's list of ways 1.13 still fails. Limits keeps that list and adds misses that showed up in later stories.

2. Where to call it

Access lists the public doors this desk has already cited: TypeSafe's API, Vercel AI Gateway, Cloudflare Workers AI, Netlify, OpenRouter, Venice, and LiteLLM. TypeSafe posted on September 20 that the waitlist is gone and that new accounts start with $5. On September 21 it posted an API and console incident, then said it was back; the status page marks Console is unavailable resolved at 08:16 UTC. On September 22 TypeSafe paused new signups and said existing accounts keep working. The post does not say when signups reopen. Price, window, and quota are whatever that host published. We did not send a request. Jevflake is a dbt package that calls TypeSafe from Snowflake SQL. duckdb-jev is a native C++ extension that batches the same calls from DuckDB. MotherDuck's prompt_jev() is a hosted SQL function on paid plans. Vercel's form router calls Jev through AI Gateway and keeps the destination when confidence is at least 0.95.

3. Who measured what

Evals collects the numbers already in the stories, with the caveat that came with each one. JevBench is a 534-decision suite that mixes accuracy with calibration, speed, and cost, so GPT-5.6 Luna can win the hard items and still sit below Jev on the composite. The table we read is v1.2.8: Jev 75.4, SemIf 74.7, djev 74.3. LangChain's judge eval puts Jev next to three LLM judges on five frozen traces. AstroHan's 30-task filter is the first keep-or-drop run on this desk that reports whether the agent still finished the task. Aman Kumar's 16,000-call write-up is the longest independent measurement we have, public cases included. Charly Poly's encoder thread is the first comparison against open zero-shot classifiers rather than a generative model. DocJev's 40-PDF pilot puts Jev next to Luna on classify and split of the same LiteParse text. RZ's 100-case security demo puts Jev next to Terra and Opus on synthetic TypeScript spans: 65.3% balanced accuracy against 75.1% and 91.2%. jevsearch's 41-query table is a ranking bench on TypeSafe's docs, 83% Hit@1 after Jev against 41% on keywords. Open-Jev's public subset scores 231 of 534 JevBench tasks: released 9B 179/231, Jev 200/231, Astra 231/231. jsort reports Pearson r 0.824 on 300 CommonLit excerpts. MotherDuck's AG News table is 89% on 100,000 training-split rows, 40 seconds, $0.50. jev-gateway's chess bench records a 57% drop in bug-fix output tokens for Astra and Sol, and a slower feature task for Opus 5. AgentOS's 121-turn chart puts an opt-in Jev router at accuracy 0.901 against its local classifier at 0.355. laya-mlx times the Laya checkpoints on an M3 Max: 13.42 ms and 7.39 ms, and matches upstream answers on 63 of 63 validation questions. Spring's Java client records a 310 ms three-question call from one laptop. Decider 1 posts accuracy 0.768 against Jev 1.13.0 at 0.727 on teacher-average labels. dejevu books the Flights search in 5.63 seconds on Groq. The Jev cell beside it, 7.09 seconds, is the other repo's published figure. A 612-row banking feed has Jev at 36.1% and Gemini 3.7 Flash at 42.5%. poker-table times a 46-spot bank at about one second median. The post's live pill says 545 ms. Bryo's 1,565 emails is a named-company table with a Gemini baseline. Compare puts those similar tests next to each other. The rows are not a ranking.

4. How people put it in a loop

Patterns groups the setups that show up more than once: rank a legal set, judge a pile of documents, override only when confident, route the expensive call, recreate the model locally. Browser Use flights is the early click-loop demo. Jevis is the same idea inside Flutter's test harness. SemIf is a 4B open rebuild that reads option logits and sits 0.7 behind Jev on JevBench. Laya is an Apache 2.0 ModernBERT-sized encoder; JevBench 70.1, and its own table trails published Jev on Banking77. Kev is a LoRA plus pointer head on Qwen that serves the same HTTP shape; author new-source table Kev-8B 79.6% versus Jev 85.7%. djev is a DiffusionGemma decision API with image input; JevBench 74.3. SimpleJev reads next-token logits from a stock open model; JevBench Qwen3.8-27B 67.3. Open-Jev is a Qwen3.5 LoRA plus a decision head. On the 231 public tasks its 9B checkpoint is 179/231. jevctl wraps Jev as shell commands, including a compact that keeps kept lines verbatim. DocJev classifies and splits PDFs with local LiteParse and hosted Jev; the author's table is 40/40 classify and 7/8 split against Luna. n8n-nodes-jev-classification is a community node with a Needs Review branch. duckdb-jev asks from SQL inside DuckDB. jevsearch re-ranks a docs search after a keyword pass. jsort orders a file from pairwise Noul judgments. The form router hands the uncertain band to Luna Fast. llamacpp-jev serves the same HTTP shape from stock llama-server, images included, with a raw label softmax. jym asks Jev only when a subcommand is missing from help text. jev-gateway asks Jev which tool a coding agent should call, and can skip the language model when every argument is closed. laya-mlx runs those Laya weights on Apple Silicon without a new accuracy table against Jev. Spring AI TypeSafe can refuse a request before the chat model runs, and can send a failed criterion back for another try. AgentOS asks Jev which model tier a turn deserves, and leaves the local classifier as the default. jev-rubiks lets code solve the cube. Jev decides whether the coach should speak, after a test where picking turns left the distance at 22. Exa's demo puts search snippets into Jev state and does not print the with-search and without-search pair. pg-jev asks from PostgreSQL and returns jsonb, with no scored set in the README. jevelry maps each answer to act, mark, or fall_back from thresholds in a JEVEL.md. poker-table asks Jev for a read of the opponent and keeps the bet math in code. dejevu runs the flights loop on a chat model and leaves its Jev backend untested. semgate is a Go HTTP middleware with a public range attached.

5. Indexes

Timeline is the same stories by article date. Limits is the failure list. Compare is similar tests side by side. People is who published them. Repos is the GitHub projects those stories cite. Terms is the glossary. Sources is the URL list. Method is how a story gets on the desk. FAQ is short answers from the same material.