EqualifyEverything / equalify-iris

The graphical-key shield added for #347 is keyed on the reason a rejection gives — "do not send it back AS INVENTED TEXT" — and the same deletion arrived under the other rule. On acir-p073 the extraction emitted each legend swatch as <dt><img alt="Dark solid fill"> States ranking in the top 12 on personal income</dt>; the verify pass answered "since the adjacent <dt> text already conveys the meaning, these decorative swatch images should use alt='' to avoid redundant announcement, per the redundancy rule in the agent contract"; the correction obeyed, and the shipped fragment is <li><img alt=""> States ranking in the top 12 on personal income</li>. The label names the category and the deleted alt named which pattern carries it, so the mapping is the half that went, while the same page's map description still names "the dotted pattern" and "the cross-hatched pattern" to a reader with nothing to attach them to. The container change in that pair is not this defect: a separate problem in the same verdict said "a <dl> with empty <dd> elements is malformed ... would be better represented as a list or a table", which is correct and is the remedy the page took. The other case is the cover: the one arm that complied with page.md's cover clause had its appearance-only alt refused because "the cover text is also transcribed in the surrounding HTML" — the transcription that clause asks for, offered as the evidence against it. So feedback.md now refuses the second door as well, and states the asymmetry without ranking itself above anything: a description that says too much costs a reader noise, one that is emptied costs them the content, the correction pass is single-shot, and page_corrected records a page that did what it was asked. The review's first round caught that an earlier draft opened "This matters more than the other notes because", which put this rule above the two clauses in the same block that keep a true report coming; the sentence now stands on its own and the pin is anchored from its capital so a comparison cannot come back quietly. Bounded by the census. Across every round log on disk 9 of 3,132 problems propose an empty alt, and 7 of the 9 are correct or harmless: three graphics omitted entirely, two existing alt="" leaving a link nameless, one doubled attribute whose empty second occurrence wins, one figure/figcaption mismatch offering "decorative" as one branch. Those stay reportable — a description may be sent back to say something different, and alt="" stays proposable where the contract itself puts the image in its decorative case. What is refused is alt="" as the remedy for a description judged redundant. Not measured after the change: spend is halted, so the evidence is the verdicts' own text and the delivered HTML on both sides of a correction, and the pins prove the clauses cannot be reworded away — not that a given model now behaves. Seven pins in test/feedback-prompt.test.ts, one mutation-checked, and the calibration doc records the recurrence beside the #347 paragraph it extends. Two review rounds, both approving; every note from the first is fixed or answered from the round log, including one arithmetic error of mine in the census enumeration. Closes #452 Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>