Leaked Jev decision-chain design: packed state plus parallel heads
A breakdown of a Jev chain that packs inbox, DOM, CRM and feed into one typed state, runs CHOICE/NOUL/SCORE heads in parallel, finishes in 3ms at $0.42 per 10,000 decisions and gates low confidence.
WHOEVER LEAKED THIS JEV LANDSCAPE SHOULD BE RUNNING SIGNAL OPS I thought agents were slow because they had too many steps to think through. Apparently not. This landscape's setup is much more interesting: → STATE packer collapses inbox, DOM, CRM and feed into one typed world, 1,008 strands indexed, no per-source branching downstream → CHOICE folder names the bucket at 562 answers per second (phishing, billing, sponsor, personal), nothing outside the labels even on retry → NOUL worth_reply fires yes-or-no at 566 per second, this youtube comment landed yes at 0.91 PASS, no "maybe later" allowed → SCORE heads (lead_quality, viral_fit, urgency) run in parallel at 581 + 607 + 672 per second against the same state, each returns a numeric with its own confidence → CHOICE next_worker closes the chain by picking the next step (review here at 0.95), the whole chain runs in 3ms, flags 7 cases where confidence crossed under the floor without ever reaching a human so you get: Sources packed. Choices labeled. Nouls binary. Chains millisecond. the interesting part isn't fewer asks per state. it's giving every ask a type the chain can trust in milliseconds. Instead of a one-model agent taking a full second to think through one bucket, you pack four sources into one state and run 8 heads in parallel at 271ms median, $0.42 per 10,000 decisions, with the gate holding anything under 0.85 at the chain. Pack the sources. Ask the heads. Chain in milliseconds.
This post quotes another post: “Long read: how to build a Jev-powered agent stack”