拆解爆款 Claude 配置：Jev 做升级路由是真的，99% 是吹的
一条号称「超越 99% Claude 用户」的配置用 Jev 决定何时升级到 Opus 5.5；作者核实 Jev 的决策数据属实，但 99% 的说法查无出处。
A VIRAL "CONFIG" PROMISES TO BEAT 99% OF CLAUDE USERS. THE TOOL IT NAMES HAS REAL NUMBERS BEHIND IT, THE CLAIM DOESN'T A post going around shares a copy-paste config that wires Claude Opus 5.5 to a five-role agent team, planner, researcher, builder, reviewer, router, with a tool called Jev deciding when to escalate. It's captioned "bookmark this now, thank me later." I couldn't verify the "upgrades your setup beyond 99% of users" line anywhere. That number belongs to the post, not to Jev. Jev itself is real. It's a decision model from TypeSafe AI, built by Diogo Almeida, a co-author of the original InstructGPT and RLHF work at OpenAI. Instead of generating text, it returns a typed decision with a probability attached. In a 500-decision test against human judgments, it matched every call, with variance 92 to 913 times lower than the other models tested, at $0.00035 per decision and about half a second of latency. The config's actual engineering idea holds up against that: use a cheap, fast decision layer to route only the ambiguous cases to the expensive model, and require a reason before anything escalates. One of its own rules: "Escalate only when evidence warrants it" The routing pattern matches how Jev is already used elsewhere, sorting research papers, scoring agent behavior. The "beyond 99%" and "thank me later" framing is the part doing the selling, and it's the part with nothing behind it. If you copy the config, keep the escalation rule. Drop the promise stapled to it.
这条是引用着另一条写的：x.com/eng_khairallah1/status/2104868464770863565