Jev: A Judgment Layer for Development

Sep 20

Jev acts as a judgment layer in the development loop, sitting between the developer's assessment of completion and handover to a maintainer, identifying potential issues and inconsistencies.

SignalJev scored that claim at 0.75 and flagged the mismatch against my recorded output.

I've been running @typesafeai's Jev inside my actual dev loop this week. Not as a code generator. As a judgment layer that sits between "I think this is done" and "I'm handing this to a maintainer." It keeps catching things I would have shipped. 1. It caught me lying in my own PR body. I'd written "551 tests passing." After a rebase the real number was 609. Jev scored that claim at 0.75 and flagged the mismatch against my recorded output. That's a stale number I would have handed a maintainer as a current fact. 2. It talked me out of my own refactor. I wanted to extract a phone matcher into a clean new module. Jev said extract, confidence 1.0, but quietly flagged scope at 0.38. I fed it the evidence that the repo's existing tests reach directly into that module's internals, and it inverted its own answer to a two line alias fix at 0.99 confidence. That was the right call. The elegant version would have forced edits to tests and docs I don't own, inside a PR with three named fixes in it. 3. It triaged a scary CI warning correctly. Green Windows E2E job, ugly cache cleanup warning. Jev called it runner cleanup rather than product failure. I checked it against the actual logs. It was right. The pattern I'd pass on: feed it the constraint evidence, not just the design options. Its first answer flagged its own doubt at 0.38 and I ignored it. Its second answer, with the real state of the repo in hand, was correct. The PR merged. Fewer bugs, and fewer confident wrong claims in my write ups, which is the failure mode I actually worry about. #Jev #AI #BuildInPublic #DevTools #VibeCoding #Omarchy

Original post