A check that reads in a third of a second can afford to read every sentence
In a test with planted errors, a fast judgment model missed one more mistake than a frontier model, but finished in a third of a second at a tiny fraction of the cost.
The AI Daily Brief (Formerly The AI Breakdown): Artificial Intelligence News and Analysis · How People Are Actually Using JevThe trap was laid in twelve passages of ordinary prose. The team at Every, which tests uses for artificial intelligence, planted seven mistakes and then set two different models to find them.
One was Jev, a model built to answer small yes or no questions about text rather than write essays or hold a conversation. The other was Claude Fable 5.1, a much larger frontier model of the kind people use for drafting and deep review.
The big model found all seven mistakes. Jev found six of seven. That near miss was the point, because Jev did the job in 0.35 seconds, compared with 8.83 seconds for Fable 5.1, and did it at about 580 times cheaper.
The result shows a shift in what editing can afford. A check that fast and cheap does not have to be saved for one careful pass over a finished draft. It can be run again and again, until checking everything costs less in time and money than checking one thing the old way.
The method requires translation. Instead of asking an assistant to please review this, a writer turns a style guide into a list of separate questions. If there are phrases to avoid, or tics that signal machine-written filler, each becomes its own yes or no question.
Those questions can then run on every paragraph of a long document. Does this paragraph contain a banned phrase, does it contain filler, does it break a house rule. Each answer is a small judgment, and together they screen for the problems a publication most wants to keep out of important messages.
Run that way, scale stops being just more of the same. A team that once settled for one generic review can now realistically check every sentence for the same flaws, and rerun the checks after every edit without thinking about the meter.
Questions people ask
What exactly is Jev?
Jev is a specialized model that makes small judgments about text, such as which category something fits, how it scores on a scale, or whether a statement is true or false. Unlike a chatbot, it does not write essays, hold conversations, or explain its reasoning.
How would I turn my style guide into questions a machine can check?
You write each rule as a single yes or no question, such as whether a paragraph contains a banned phrase. The model then answers that same question for every paragraph, returning a probability from zero to one, so you can block or flag anything above a threshold you set.
What does 580 times cheaper actually mean in practice?
It means the cost of computing the answers is roughly one five-hundred-eightieth as large, so you could run the same screen hundreds of times and still spend less than one review by the larger model. In other tests described in the episode, large screening jobs cost cents, not dollars.
Could this replace a human editor?
No. The test shows Jev can catch many routine violations quickly, but it missed one planted mistake that the larger model caught, and it gives a score without an explanation. It works best as a first screen for every sentence, with people or larger models handling close calls.