用 Jev 给 Obsidian 笔记库做类型化校验
作者给 wiki linter 加了 Jev 层，对 96 页笔记跑 849 个类型化问题，约 95 秒、0.07 美元，查出 9 处引用来源不含所引论断。
added a jev tier to my wiki linter this week. 96 page obsidian style vault. jev is typesafe's small typed judge model, you send it text plus typed questions and it answers, it does not write anything. one run, 849 typed questions: > page vs frontmatter drift, as a score > contradictions between related pages, yes or no > for every evidence line, does the cited raw source contain the claim, in 24k character chunks > about 95 s > 1.69M tokens > about $0.07 > 0 drift, 0 contradictions > 9 evidence lines whose cited source does not contain the claim > spot checked one of the 9, it was real the linter never edits the vault. it prints the list and i read it. did not use it for: a skill router, a browser layer, a game macro. agents did most of the implementation.