Jev typed decisions cut LLM costs 350x, CMU researchers say
The author routes model selection, next actions, risk blocking and completion checks through Jev's Choice, letting Claude only write text, and claims 10,000 decisions cost just $0.42.
nobody is talking about this and it's insane A Carnegie Mellon University researchers have found a way to use JEV to make your LLM run 24/7 at a lower cost in x350 the LLM only does what needs words. everything else is a typed decision: > which model? Jev routes every task to a tier before a single token is spent > who acts next? one Choice picks the next worker keep or drop? every tool output checked, the noise dropped > risky? anything that sends, pays or deletes gets flagged before it runs > done? Jev checks the goal is actually met, not just "looks finished" > Claude writes research notes and drafts, the only place in the whole folder where text gets generated hard rules: stop after ten actions, never publish anything unapproved > receipts: options offered, chosen id, re-check, fallback, every call logged > the thresholds come straight from my own evals: act only on confident answers, fraud needs 0.95, tuned on dozens of labelled traces and the fees are the whole point > Jev: $0.42 for 10,000 decisions > median: 300 ms per call a whole day of decisions on a frontier model costs dollars. the same day on Jev costs cents you stop babysitting the agent. you read the receipts
This post quotes another post: x.com/thegreatest_sv/status/2102802378378621355