Claude proposes, Jev decides: cutting 1,000 decisions from $605 to $0.17
Claude drafts plans, Jev picks from options prepared by code, and code validates the final choice, with every step leaving a replayable receipt; 49 of 49 WebMCP tasks passed.
CLAUDE + JEV ENGINEERING = THE HARNESS SKILL THAT MAKES AI ENGINEERS HARD TO IGNORE IN 2026 in this setup, 1,000 decisions on a cold frontier model cost $605. route the same 1,000 through the harness, and the bill drops to $0.17 the core idea: Claude proposes, Jev answers, code decides every step leaves a receipt the loop: > propose - Claude drafts plans and patches; execution authority stays with code > filter - code removes routes your machine can't run before Jev sees the menu > answer - Jev selects from the prepared options with a probability, or abstains > re-check - code validates the choice against live state. stale choices go to fallback > act - the tool runs through the normal approval path > receipt - every step is logged and can be replayed against a different policy the numbers: > 0.2s median per call > ~$0.0002 per desktop step > 49 of 49 WebMCP tasks passed the part worth paying attention to: "I don't know" is a valid answer when Jev is only 38% sure, a fallback defined in advance takes over instead of forcing a guess worth stealing even if you never use Jev: > a decision the harness can check the exact options code prepared > a defined fallback when confidence is too low > a way to measure what actually happened walk into a meeting, turn a $605 line item into 17 cents, and people remember who should design the next system. Claude writes the proposals, Jev judges the options, and code keeps the final say
This post quotes another post: x.com/polydao/status/2103689373774483815