56 transcript sections went to JEV. 30 requests came back with answers and 26 hit overload errors. Not one section met the rule for safely discarding it, so all 56 still needed review. Before that test I work out what a classifier does with ChatGPT, Claude and Qwen, and where JEV sits next to a chat model.

What JEV does

JEV from TypeSafe AI is a classifier, not a chat model. You hand it a schema of allowed answers and it returns one of them with a probability score. A paragraph of free text is not an option it can pick.

JEV doesn't search and it doesn't clean data. Filtering a raw database is faster and free with normal code. The data has to be prepared first, by me or by a cheaper model, before JEV sees it.

  • Which category a piece of text belongs in
  • How severe something is
  • A yes or no answer with a probability

The liquor store example

Say I want to sell tequila. I download a public database of liquor stores in Nebraska and give JEV the exact criteria for the stores I want. JEV sorts the clear cases and a stronger model researches the rest.

The payoff shows up at scale. On a hundred rows JEV doesn't add much. A plain filter already answers the question when the answer sits in a column.

  • Clean the messy data by hand
  • Run a cheaper model over it first, then pass it to JEV
  • Pay more and let a stronger model do the whole job

Testing it on my own edit

I pointed JEV at my own transcript to find sections worth cutting. Judging sections means defining the test first. To ask whether a section has a strong hook, I have to spell out what a strong hook is.

56 sections went out. 30 requests returned answers and 26 hit overload errors. None met the rule for safely discarding a section, so every one of the 56 still needed review. The review workload stayed the same.

JEV ran free on the Vercel Gateway until September 25, and TypeSafe gave me a $5 credit balance. Both belong to the September 20, 2026 recording. The 30 successful responses recorded $0 in gateway charges. Failed requests returned no billing metadata, so no cash saving was established.

Claude and Qwen

Claude stated wrong things as fact. My payment method sits in my Vercel billing settings. Claude missed it until I opened the page and pointed at it. Then Claude claimed I had set up a TypeSafe account for ChatGPT under the same promotion. I never set one up.

Qwen reported that JEV has no paper and no published weights. Qwen also reported that the training method for calibrated probabilities is described only in a blog post. A reproduction built on a Qwen base appeared within hours of the launch.

I call that reproduction Chinese JEV while reacting to Qwen. I didn't establish the author's nationality or that the reproduction matches JEV's capabilities. The page I opened belongs to Harsha Gundala, and Qwen's answer is part of the recording, not independent verification.

Tools and links

  • TypeSafe AItypesafe.aiThe company behind JEV, where my account and $5 credit sit.
  • JEV documentationdocs.typesafe.aiThe official reference for JEV.
  • TypeSafe launch post by Diogo Almeidatypesafe.aiThe launch post for JEV and the qualifications behind it.
  • Vercelvercel.comIts gateway is where I ran JEV during the free launch promotion.
  • ChatGPTchatgpt.comI asked it to explain JEV in simple terms with exact examples.
  • Claudeclaude.aiI worked through the healthcare provider data problem with it.
  • QwenI asked it what TypeSafe claims about JEV.
  • Qwen-based JEV reproduction on Hugging Facehuggingface.coHarsha Gundala's open reproduction, the page I opened after Qwen's answer.