Every ran 777 judgments in under a second and still wanted a better accuracy check

Mike Taylor, head of evals at Every, published a launch-day write-up of Jev on September 15, 2026, and updated it on September 18.

He fed Jev the text of his 27 Every articles plus 10 AI-styled counterparts, then asked the same 21 questions of every document. The questions came from Every’s list of AI writing tells: repeating an idea without adding evidence, forcing a symmetrical both-sides argument, overexplaining a straightforward point.

Jev returned 777 answers in under 0.7 seconds. Taylor estimated the cost at a quarter of a cent. He wrote that the model correctly flagged pieces of his that leaned more on AI, and that he would want a more thorough accuracy check before putting it into production. He called that useful as an early warning. The alternative, he wrote, is not checking at all.

He ran 11 experiments in total, covering context lookup, work checks, and routing decisions. Across that suite Jev produced 1,709 judgments for an estimated total of less than a cent. GPT-6 Astra, running in Codex, wrote the queries from TypeSafe’s docs.

Every CEO Dan Shipper ran a tighter test: 12 synthetic passages, six clean and six with planted problems, scored on four writing checks. Jev’s median time was 0.35 seconds per passage. Claude Fable 5.1 at high effort took 8.83 seconds, about 25 times slower and about 580 times more expensive on Taylor’s estimate. Fable caught all seven intended defects. Jev caught six, and missed the same one in all three runs: “a shared appointment calendar that parents and staff teach together,” which Fable treated as an unexplained action.

Taylor’s advice was to pick a judgment you already make repeatedly, compare accuracy against the model you use today, and see whether acting on the answers makes the work better.