EqualifyEverything / benchmarks-ai-alt

Nine batches, 30 items each except the last. 78 ready, 172 dropped: 143 because the shipped alt is not good enough, 21 because the alt is good but the item is not functional or duplicates one already kept, 8 because the alt states something untrue. review/batch-01-notes.md sets the judgment standard and each later notes file records where that standard was refined and which earlier verdict the refinement contradicts. apply-review.mjs refuses to re-review a decided item by design, so the contradictions are written down rather than quietly fixed. review/RUN-2026-08-27.md reports the real counts and the harvester bugs the review surfaced, ordered by how much fixing each would improve the next run. The validation tool now shows the accessible name rather than only the alt attribute. 36 of the 78 exported items carry no usable alt on the image itself, because the text lives on the link or button, so the old display showed "(empty alt text)" and nothing to judge for nearly half the corpus. The alt attribute moves to the context list, distinguishing a missing attribute from a deliberate alt="". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>