Jev sources: where every figure here comes from

Everything on jev-agent.com traces to a public source or to a run you can repeat. These are the Jev sources behind each kind of figure, how our own numbers were measured, and every correction we have published against ourselves. The official documentation was last checked 2026-09-19; project READMEs and JevBench on 27 September 2026.

How we handle facts

Technical behaviour (primitives, request shape, SDKs) comes from the official docs.
Pricing, model routes and context limits are cross-checked against provider listings.
Company and funding details come from launch coverage.
Third-party results are quoted with their author, date and caveats, and linked; we do not re-run someone else's benchmark and publish it as ours.
Our own measurements say when, how and through which route they were taken.
Anything we cannot source, we leave out rather than guess.

Jev sources for each kind of figure

Figure	Source	Shown on
Price: $0.042 per million input tokens, nothing for output	TypeSafe's documentation, cross-checked against the OpenRouter and Vercel AI Gateway listings	pricing, plans
Latency of 70–500 ms, the 64k and 32k token budgets, 255 options, rate limits	TypeSafe's HTTP API and models reference	API reference, limitations
Versions and aliases: jev-latest, jev-preview, jev-1.13.0	The models reference, plus our own calls: the models list shows only the aliases, yet the versioned ID is accepted	API reference
Failure modes: counting, dates, indirection, adversarial text	TypeSafe's jaggedness page for jev-1.13	limitations, System One
Funding, co-founder and launch date	TypeSafe's launch post and launch coverage	TypeSafe AI
The vendor accuracy table (67.8% against 74.1%) and the speed and cost multiples	TypeSafe's launch post; its accuracy means agreement with two frontier models, not ground truth	Jev vs LLM, limitations
The Doom bot at about $7 an hour, and Wikiracing	TypeSafe's launch post	pricing, Wikirace
Phishing accuracy: 62.6%, 81.3%, 95.1% and the controls	anisselbd/jev-phishing-bench	Jev vs BERT, phishing
JevBench scores for Jev and the open models	JevBench v1.4.2.1, published 2026-09-27	open source
Open-source projects and showcase entries	Each project's own README, opened and dated when it was added	open source, showcase
Latency, cost, accuracy and game results measured on this site	Our own runs; the harness and data are in jev-measured	benchmarks, use cases

How our own numbers were measured

The eight use-case calls were each run three times through OpenRouter on 18 September 2026, with the network round trip to its host measured alongside: about 200 ms of every latency figure on those pages is that trip. Token counts are the ones the call was billed for.

The head-to-head put the same state and questions to Jev and three chat models, five runs per fixture, on 20 September. The accuracy set is 27 tickets with exactly one decisive signal each, plus six deliberately ambiguous ones with no label, run three times. The game benchmark played 100 TextArena games per game against a bot that always takes the first legal move. Each page gives its own method and limits, and says plainly when a sample is too small to settle a question.

What does not count as one of our Jev sources

Some things look like Jev sources and are not. A vendor number without its method is quoted as the vendor's, beside the method, and never restated as our own finding. A post or screenshot that cannot be opened again is left out. So is a chatbot's answer about Jev: language models repeat launch-week guesses about pricing and access with complete confidence. And a page that errors once is not proof that something is gone. OpenRouter's Jev listing once answered a script with a 404 while loading normally in a browser, so a missing page is checked by hand before we call it missing.

Corrections we have published

A site that measures things gets some of them wrong. These are the corrections to our own figures and code, newest first. The pages were changed rather than footnoted, except where the original result is itself worth reading.

2026-09-27The phishing study was quoted at 95.1% without the controls its author added: a two-line regex reached 91.6%, and Claude Haiku 4.5 asked the same five questions tied with Jev on held-out mail. Every page that cites the study now carries both.
2026-09-27The fixture mean was given as 440 input tokens; the measurement file says 447, and the support ticket is 539. An 8,000-character state was said to cost eight credits; in English it is about 2,000 tokens, so two.
2026-09-27The documentation listed running out of credits as a 429. The API returns 402, and never 429.
2026-09-27Code samples compared a Score result to a label string. A Score is a number from 0 to n−1, so those checks could never match; they now threshold the rounded score.
2026-09-27The game benchmark said its harness was in the public jev-measured repository. It is not yet, and the page now says so.
2026-09-23Pack sizes were translated into requests at 440 tokens each, which overstated them 2.3 times: every request under 1,000 tokens costs a whole credit. The plans page now says the credit count is the request count.
2026-09-20We reported that chat models returned the wrong types. That was our request body: structured output was off. With it on, every model returned the right types on every run. The correction stays published beside the original table.
2026-09-20“Chat models are inconsistent, Jev is not” was drafted before we checked it. At default sampling the chat models varied between identical passes; at temperature 0 none did. The claim was dropped.
2026-09-18The first code samples called client.systemone() and shaped the questions from the prose of the docs. The SDK method is system_one and the questions are typed objects; every sample was rewritten against the real package.

Primary Jev sources

Report a wrong figure

Jev is weeks old and changing quickly, so something here will go out of date before we notice. Email support@jev-agent.com with the page, the figure and where you saw a different one. One person reads it. A confirmed error is fixed on the page, re-dated, and added to the corrections above.