DiffJury answers nine PR questions in one Jev call

rvaniaaa @rvaniaaaaOriginally posted 2026.10.09

Open original on X

This is f*cking insane. One Jev call answers nine questions about a pull request, verdict included DiffJury's nine, sorted by question type: → Score, how much: risk, missing tests, docs debt, blast radius → Noul, yes or no with a probability: needs design review, needs security review, merge blocker → Choice, one from your list: review depth (skim, standard or deep) and the verdict (approve with nits, request changes or block) DiffJury is a demo of the format. any decision you settle by gut can be written the same way, and the three tests below decide whether it belongs on Jev how to pick the tool for your own question: → "which one of these?" is a Choice, like review depth, and the options are yours → "how much?" is a Score, like risk, and you write what each level means → "yes or no?" is a Noul, like needs security review, and you get a probability instead of a verdict ask all of them in one call about the same state. killmyidea does it with ten questions, then plain code turns the answers into kill, fix or ship with fixed thresholds every answer comes with a confidence, and a high one means Jev is sure, which is a separate thing from being right. set a floor and send anything under it to the big model before you move a question over, run three tests: every valid answer can be listed in advance, a person would settle it at a glance, it happens often. fail one and it stays on the big model your first run: → take one decision you settle by gut, write it as a few questions, and label each Choice, Score or Noul → run 20 old cases through and compare the answers with what you decided → check how confident Jev was on the ones it got wrong: if it was unsure, that's where your floor goes. if it was sure and wrong, that question stays on the big model full Jev guide below

This post quotes another post: x.com/rvaniaaaa/status/2101709732314558639