Graph engineering plus Jev sorts 1,000 articles for 4 cents
A workflow method that hands every decision arrow to Jev for choice, score or probability, letting big models process only the survivors.
THIS IS CHEATING. graph engineering + Jev engineering = agents that work in parallel and decide in milliseconds. 1,000 articles sorted overnight for about 4 cents. people already ran Jev on 1,018 research papers for 8 cents. the setup: 1. Draw it as boxes and arrows. boxes are the work: read an article, summarize it, write the final report. arrows carry results forward. then ask of every arrow: does the next box need this result? if not, the arrow is fake, and both boxes can run at the same time. in a 40-step workflow, a dozen fake arrows is realistic. 2. Mark every arrow where something has to decide. is this article worth keeping? is this summary safe to publish? retry or stop? 3. Give the boxes to your big models. they read, summarize and write the report. 4. Give the arrows to Jev. it answers in one of three ways: pick one option, score on a scale, or give a yes/no probability. about 100 ms per answer, $0.042 per million tokens. “keep or drop” on 1,000 articles is roughly 4 cents. the same 1,000 calls on a top model is roughly $4. 5. Let plain code follow the answer. Jev drops the weak articles. say 975 are gone. the other 25 go on to the big model for the summary. any answer Jev isn’t sure about goes there too. your code picks every branch. two rules: 1. a high confidence from Jev can still be wrong. set a confidence cutoff for each decision. 2. one question per decision. Jev can’t chain one judgment into the next. keep one check no model can argue with, like a test that ran or a number that landed. start with the one arrow that fires most often. run 20 items, read the bill, then add the next.
This post quotes another post: x.com/imryven/status/2101412287013593420