EqualifyEverything / benchmarks-ai-alt

First real round, run in pi with curl and no web search. Worked from the WAI functional images tutorial plus gov.uk, loc.gov, harvard.edu, nps.gov, gutenberg.org and geology.com. Dropped BBC and MDN as client-rendered, which is the right call: markup that only exists after script runs cannot be verified from what the server sent. Reviewer rejected 5 for leakage and sent 4 back for provenance that did not match the page. That is the separation working as designed on its first outing. Committed as the audit trail, rejections included. A rejected item is evidence about what does not belong in the corpus, and a corpus whose construction cannot be audited cannot be defended. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>