EqualifyEverything / benchmarks-ai-alt

The seeking directive knew what to verify and what to write, but its finding section was six abstract strategy categories. An agent given that will collect whatever it stumbles into, which is header logos and footer social icons: the two shapes that fill fastest, duplicate worst, and require the least judgment. - A hunting recipe per sub-type, all seven, naming where it lives, the markup signature that identifies it, and the judgment that makes such an item worth having. Including which shapes to prefer, because the easy instance of a sub-type usually cannot discriminate between models. - Extraction mechanics for a harness whose only tool is a shell: curl, the signatures to grep for, client-rendered pages that cannot be verified from served markup, and politeness including a robots.txt check. - A section for harnesses with no web search, since pi has none, with stable named starting points and an instruction to record that search was absent so the reviewer reads a narrow round correctly. - Deliberate hunting for the three difficulty quotas. Empty-alt, dual-purpose and ambiguous items do not turn up on their own: the empty-alt signature is the opposite of the icon-only shape the rest of this work chases. - Concrete near-duplicate detection: file name stem, sprite symbol id, icon font class, SVG path data. Two sites on one component library ship identical markup, and that is a duplicate however different the domains are. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>