Building a 24/7 HFT bot on Jev with six structured checks to filter false signals
The author built an HFT bot on Jev's Choice and Score primitives that filtered 3 real signals out of 1,000 generated in one night, without generating a single word.
I built a 24/7 HFT bot on Jev. It generates 1,000 trade signals a night. 997 of them are lies. Jev catches the 3 that aren't - without writing a single word. Six months ago one of those lies got through. Cost me 4% in a single session. At this speed there's no catching it after - the window closes before a language model finishes its first sentence. So the bot doesn't use language. Not for any of it. Six checks. Each one fast enough to beat the window. Break any one, the trade never sees a dollar. 1. Choice first: trade it, shelve it, kill it. Jev scores all three at once - if the top option isn't ahead by a real margin, it's a shelf, not a trade. 2.Score next, never a label. Conviction lives on a line from 0 to 1. A 0.51 and a 0.94 are not both "yes" - they get sized nothing alike. 3. One Noul question before anything else moves: does this overlap what I already hold. Above my correlation threshold, it dies here, no matter how good the idea looks. 4. Confidence sets size, not me. Above 0.9, full Kelly-scaled risk. 0.6 to 0.9, half. Below 0.6, zero capital, no exceptions. 5. Every number comes calibrated. Jev is trained under RLCD, so a 90% answer is right about 90% of the time - not confident-sounding, actually correct. 6. Nothing in the stack writes a sentence, ever. Typed answers only, in the low milliseconds. The idea can still be wrong. It cannot come back malformed, and it cannot talk its way past a check. The rule that survived every version before this one: the model that dreams up the idea never gets a vote on whether it's safe. Wrote the whole system up - every threshold, every primitive, the full architecture. Paper below.
This post quotes another post: “Running a nine-agent hedge fund with GPT-6 Astra”