WHAT NEW IN X10.30 FRI

X 上的 AI 合集档案 · 22 个合集 · 60,637 条

jevVIDEO

Jev 三级瀑布过滤百万条市场头条

用 Jev 先给每条头条打概率,只把不确定的交给 Haiku、分歧的交给 Opus,成本从 800 美元降到约 89 美元,并用 Opus 结果反调阈值。

xinces
@xinc3s8 天前收录原推发布于 2026.10.20
xinces @xinc3s原推发布于 2026.10.20
在 X 上打开原推

1,000,000 MARKET HEADLINES A DAY. OPUS 5.5 READING ALL OF THEM: ~$800. THE SAME DAY THROUGH A 3-MODEL CASCADE: $89. AND OPUS ONLY SEES 1% OF IT the trick isn't a smarter model. it's a waterfall instead of: every headline → opus → $800 it's: every headline → jev → only the unsure ones → haiku 5.5 → only the fights → opus 5.5 jev can't write a single word. it returns a pick and a probability. that number is the whole filter: → p ≥ 0.90: done. dropped or flagged, 0 tokens out → below 0.90: haiku 5.5 reads it and writes a card with a source link → haiku and jev disagree: opus 5.5 gets it. about 1 in 100 one day, list prices, my math: jev: 1M × 100 tokens × $0.042/M = $4.20 haiku 5.5: 80k unsure ones = $5.20 opus 5.5: 10k hard ones = $80 total $89.40. opus on everything: $800 the part nobody does: that 1% opus sees is a free answer key every sunday, count how often opus overruled jev. more than 1 in 5? raise the cutoff. less than 1 in 20? lower it. the filter tunes itself on its own mistakes where it pays on a market desk: → headlines: noise or material? → 8-Ks: routine, or the one that moves guidance → form 4s: planned sale or a real insider buy → earnings calls: same script or one changed word → watchlist alerts: 300 a day → the 6 worth opening the 3 jev questions i start with: action: choice → drop / flag / unsure material: noul → "this changes revenue, guidance or risk for [ticker]" insider: choice → planned 10b5-1 sale / open-market buy / other paste this into claude code ↓ "Build a 3-tier headline cascade in this folder. Watchlist: [tickers]. 1. http://ingest.py: pull new headlines, 8-Ks and Form 4s for the watchlist into inbox/. One JSON per item, keep the source link. 2. http://gate.py: send every item to Jev (typesafe-sdk) with the action and material questions. p ≥ 0.90 → done and logged. Cutoffs live in config.yaml, never in a prompt. 3. Everything under 0.90 → subagent reader, model haiku: one card, the number, the source link. No source = "unknown". 4. Card disagrees with Jev → subagent judge, model opus, read-only: one-line verdict + why. Never sees the price. 5. Log every Opus overrule to overrules.csv. Every Sunday print the overrule rate per cutoff. 6. Read-only. No brokerage. Show me the files first. Nothing runs until I say go." you don't need a smarter model. you need one that knows when it isn't sure

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

x.com/xinc3s/status/2112514077754863940采集于 2026.10.21 10:03