Haiku 读新闻、Jev 决策、Opus 下单的炒股流水线

xinces @xinc3s原推发布于 2026.10.11

在 X 上打开原推

I GAVE HAIKU 5.5, OPUS 5.5 AND JEV MY BROKERAGE ACCOUNT ON THURSDAY MORNING. BY FRIDAY’S CLOSE THEY’D MADE $36,015. I DIDN’T PLACE A SINGLE ORDER the best trade happened at 6:12 pm, after the market closed. the smartest one was a trade they refused who does what: → HAIKU 5.5 reads everything. headlines, filings, press releases, 24/7, for pennies. it writes one card per story: what happened, the number, the source, the minute it saw it. no opinions → JEV decides. it can’t write a single word. it reads each card and says buy, sell, short or skip, with a probability. under 0.80, nothing happens → OPUS 5.5 trades. it picks the size, the entry and the exit, and sends the order. but first it has to write one sentence: why this news moves this price. no sentence, no trade → me: i set the limits on wednesday night and went to sleep thursday: +$11,000 7:06 am, pepsico beats. haiku had the card in 41 seconds. jev: buy, 0.86. opus went long $80k at 7:09. +$2,480 10:37, goldman upgrades palantir. long $60k. +$1,560 midday, the ft reports starbucks looked at buying chipotle. long $90k. +$4,860 2:39 pm, the ft says openai’s revenue is $20B below what it signaled. jev: sell, 0.91. opus dumped my nvda before the chips slid, then shorted the chip etf. +$2,100 6:12 pm, spacex buys spectrum to turn starlink into a phone carrier. carriers get crushed after hours. opus shorted t-mobile with $100k before i finished dinner. friday it closed −13%. +$11,200 friday: +$25,250 8:07 am, delta misses and cuts guidance. every human instinct says buy the dip. jev: skip, 0.34. delta closed flat. no money stuck in a dead trade 8:31, humana’s medicare ratings come in better. long $110k premarket. it ran 12%. +$11,440 10:05, tower stocks rip 7 to 16%. jev said buy. opus refused: it couldn’t explain why towers go up when spacex just found a way around them. no sentence, no trade. that one would’ve paid. i’m fine with it 11:20, big tech bounces. long amazon $90k. +$2,610 gross $36,250. fees and borrow $171. models $64. net $36,015 what actually got me: the biggest trade came from a headline at 6:12 pm, when every human trader was already offline and the most expensive model never traded on a headline it couldn’t explain the edge isn’t a smarter model. it’s three simple jobs and one rule: can’t explain it, don’t trade it the prompts i used ↓ HAIKU 5.5 “watch the news for [watchlist] 24/7: press releases, sec filings, ft, reuters, wsj, x. for every new item write one card: ticker, what happened in one line, the number, the source link, the minute you saw it. no opinions, no predictions. no source = ‘unknown’. never skip a card because it looks small.” JEV (not a prompt, 3 typed questions) action: choice → buy / sell / short / skip priced_in: noul → “the market already moved on this news” size: score → small / normal / large thresholds live in code: act only above 0.80, skip if priced_in > 0.70 OPUS 5.5 “you only get cards jev cleared. before any order, write one sentence: why this news moves this price, and what would prove you wrong. if you can’t write it, don’t trade. max 30% of the account per position, max 4 open at once. every position gets a −3% stop. shorts only on liquid large caps. no options. log every order with the card that triggered it.” ME (limits in the broker, not in a prompt) api key with trading on, withdrawals off. daily loss cap $8k: everything closes and the agents stop. kill switch on my phone.

这条是引用着另一条写的：x.com/xinc3s/status/2108204467774509059