量化台用 Jev 前置过滤 430 万次交易决策
50 个永续合约每秒约 50 次决策，Jev 先答噪声/regime/是否值得交易/仓位四问，低于 0.85 直接丢弃，只有 1/70 调用到 Claude，日成本从 4.3 万降到约 800 美元。
4.3 million trading decisions, only $800 that's what happens when a quant desk puts jev in front of claude 50 perps, every tick, 24/7 ~50 decisions a second, 4.3 million a day send all of that to a frontier model and you pay about a cent per call that's $43,000 a day for a bot that mostly says hold jev charges $0.042 per million tokens and nothing for output so the desk flipped the order > is this noise or a real move > what regime is the market in > is the setup worth a trade > how much size does it deserve jev answers all four in a fraction of a second anything under 0.85 confidence dies in the router only 1 in 70 calls ever reaches claude jev ~$180 a day claude ~$620 a day total ~$800 the cheapest model in the stack now makes 98.6% of the calls and the expensive one finally only thinks about trades worth thinking about by the time a frontier model finishes reasoning on every tick, the candle has already moved jev decides before it does
这条是引用着另一条写的：「用 Jev 从零搭高频交易 bot：每区块一次买卖决策」