Claude 提议 + Jev 决策：把 1000 次决策从 605 美元压到 0.17 美元
用 Claude 起草方案、Jev 从代码准备好的选项里选、代码做最终校验，每步留可回放凭据，WebMCP 49/49 通过。
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
这条是引用着另一条写的：x.com/polydao/status/2103689373774483815