Jev handles decisions while Haiku 5 only writes code
Jev runs 31 typed questions in 0.4s and picks 4 files out of 260 while Haiku 5 writes the patch, shipping in 3m 40s for $0.07.
JEV + HAIKU 5 is outrunning my frontier stack... I took every decision away from the model. It only writes now. Hand it a task → Jev decides what it sees, where it goes, whether it runs → the model types. Full production ship in 3m 40s: → Jev scores the request before anything loads: answerable, in scope, not already done → 31 typed questions run in one call. 0.4s total → It picks the 4 files worth loading out of 260 → Haiku 5 writes the patch. That is its only job on this run → First patch failed the test. Jev caught it, re-scoped, second one passed → Jev holds the commit until the diff clears the permit rule 31 decisions. 2 model calls. 0 prompts rewritten. The bill: → Frontier stack, same task: $4.18 → This run: $0.07 → Not because the model is cheaper. Because it ran twice instead of nine times → A typed question costs less than the whitespace in your prompt Prompt engineering was people guessing what to say to the model. Jev engineering is deciding what the model never has to see. Not a router. A router picks a model. Jev picks whether there is anything to send. Jev decides, the model writes. Nobody needs a bigger model for the deciding part.
This post quotes another post: x.com/0xcodio/status/2104195432209387524