Trading assistant rebuilt with four seats and a five-step cap
The author applied the compounding-error data from Dots' system card to their trading bot, using four roles (LOOKOUT, MAKER, SKEPTIC, RUNNER) and a five-step checkpoint to filter out 326 suspicious pools overnight.
this is f*cking gold, the best thing i read this month... openai shipped dots and everyone posted the plush mascots. one person posted two numbers from the system card instead and wrote a 4,000 word guide. the case is simple. a chain of 5 steps produces 8.6% flagged problems. a chain of 10: 19.7%. errors don't add up. they compound. a step that drifts slightly hands a wrong picture to the next one, and the next one builds on top of it. for memecoins this is death. the bot found a token, checked holders, calculated concentration, evaluated momentum, decided to trade. if step 1 caught a scam, every following step builds a case for buying that scam. the fix: one agent, four seats. LOOKOUT only monitors. MAKER analyzes. SKEPTIC checks and finds what is wrong. RUNNER executes after approval. one memory, one rulebook, four narrow jobs. and one rule that changes everything: max 5 steps. after the fifth SKEPTIC runs a checkpoint. fail = hold. nobody moves. 98.6% of tokens on http://pump.fun are scams. the seat that finds problems does more work than the seat that finds opportunities. i applied this to my bots. overnight: 340 pools. SKEPTIC killed 326. 14 reached MAKER. 3 filled. +2 SOL not because the bots found good tokens. because SKEPTIC killed the bad ones before MAKER fell in love with them. the entire article in one line: don't build a longer chain. build shorter legs with a skeptic at the end of each one ↓
This post quotes another post: x.com/leanxbt/status/2106055029207568824