第二大脑：GPT-6 Astra 写笔记，Jev 只做守门判断
用 Jev 分两轮判断「先选文件、再选段落」，检索准确率从 40% 提到 96%，每问 token 从 2.1 万降到 1200。
GPT-6 Astra writes the vault, and Jev is the guard that stops the rot before a single sentence gets saved. Clip labels the split in the first frame. Astra writes. Jev stops the rot. That is the same cut as the second brain guide: one model builds the notes, the other only judges. Square is the free pass. One block, code checks it, no model call. Expired pages, duplicate text, and anything over budget die here, so Jev never sees them. Unfold is the shortlist. Panels leave the square and become a body you can inspect. In the vault that body is a capped set of candidate notes, not the whole folder. Robot frame is Jev on guard. One call, every pair. It never writes a line. It only returns a label you already defined: duplicate, related, revises, contradicts, or none. Fold puts the same parts back into one object. Link proposals sit in a review queue. Notes stay plain markdown. Nothing edits itself. Launch is the rule on the last card. Astra writes last, or the vault does not know. Recall runs the same way: which files, which sections, enough evidence, then the writer. Skip Jev and Astra guesses a file by title, reads all of it, and hopes. On 52 questions that found the right section 40% of the time, at about 21,000 tokens a question. Two Jev rounds, files then sections, moved that to 96% and about 1,200 tokens. Same agent, same notes. Only the decisions changed. Ask for a fact that was never written, like a favorite pizza topping, and the old path returns an unrelated file. Jev scores it low and the vault says it does not know. GPT-6 Astra still does the interview, the compiled pages, and the final answer. Jev only stands at the gates. Writer last is what keeps a growing folder from rotting in silence.
这条是引用着另一条写的：x.com/rewind02/status/2104888148568277164