EqualifyEverything / equalify-iris

#334's marked run-in-run axis: a printed marker that ADVANCES makes a run of bold labels a list, and the rule is written on every branch that reads the shape. Three of the page prompt's rules reached one printed shape — a run of paragraphs opening in bold type — and none of them said which wins. LISTS ("Typography does not decide this"), NAMED ITEMS AND THEIR EXPLANATIONS (a <dl>), and the run-in heading rule are each present, specific and individually right. acir-p059 prints two devices in that one italic typography: "Property taxes." (unmarked, a sub-topic) and "a. Locally assessed land and structures other than public utilities." (marked, one of a lettered run continuing on the next page). Measured over the round logs already on disk, all $0 and read-only: 68 draws carrying at least three of each run's labels, over three printed runs and five page files. 12 put the run in a list. 36 landed the label in a <p> — 34 of those opening it in emphasis, which is the shape NAMED ITEMS names verbatim and rejects. 14 emitted a heading run and 6 a <dl>. So the plurality answer is the one no rule allows, and the 20 draws that did pick a member of the collision picked two different members. That spread is what an undecided precedence looks like. Within-reply control: on acir-p060, 22 of 22 of those same replies emitted an <ol> or <ul> for other content on the same page, against a base rate of 43.1% of 3,525 replies over the other 258 pages. The tie-break and its reason go under NUMBERS THE PAGE SHOWS, where the marker's own rules live: the sequence is the page saying these items belong together and in what order, and no other element carries it — a heading run and a <dl> both have somewhere to put the name and nowhere to put the letter. It names the answer it overturns (a marked run stays a list even where its items run long, the one place NAMED ITEMS' substantial-content test no longer sends you to headings) and four failure paths: one marked paragraph is not a run; a marker that repeats rather than advances is not a sequence; a partly marked run is still one list, logged; and a run continuing from an unseen page carries start. Because a rule binds only where it is read, the exception is cross-referenced from all THREE other branches that reach the shape — NAMED ITEMS, the sub-topic promotion rule, and the heading-LEVELLING clause. That third one was found by review, not by me: its own example is a step label "B.", and what it does is assign a level, so it never reads like a rule about what the shape is. It also reaches the table-of-contents group-label clause beside it, which is kept and named rather than scoped away — an <h3> in a contents page holds a printed letter no better than any other heading — with the entries of a marked group nesting inside its own <li>, because one flat list carrying both asserts they are peers. CONFOUND, recorded because it decides how much of this to believe: all 68 draws read an agents/page.md blob predating #432's `type` clause, the clause a lettered run most needs. The two lettered subjects are therefore latent-or-fixed and only a paid round can say which. Digits are the part that indicts the prompt as it stands: `start` was already specified in the blobs those draws read, so #432 changed nothing for them, and 16 of 27 still came back outside a list. Two of my own figures were wrong on the way here and both are corrected in the body rather than dropped. "0 of 59 in a list" keyed the denominator on the printed marker, which a correct <ol type="a"> drops — so every draw that got it right fell out of the population. "56 of 68 emitted none of the three" was the complement of one outcome in a four-outcome population; not-in-a-list contains the <dl> and heading answers, which are two of the three. The real figure is 36. Two axes are refused with figures so they are not rebuilt: a "≥2 run-in labels, 0 heading elements" detector fires on 46% against a 48% base rate, its 25 label signatures splitting into at least five hand-checkable classes that co-occur on one page; and 425 of 2,634 replies with no <img> have a log naming a picture across 133 pages, where the matched text is the CORRECT answer for a decorative cover described in words. Prompt text only, mirrored in both copies, no code path touched. Four review rounds, each round's finding created by the last round's fix. npm test 1650 pass / 0 fail — unchanged, because pins added inside an existing assertion table are not new test cases, so every pin was verified by mutating its clause and confirming that pin's own message fails under node --test. #334 STAYS OPEN: the visible U+002D line-break hyphen axis and the dropped cover are untouched. Refs #334. Co-Authored-By: bbertucc <46652+bbertucc@users.noreply.github.com> Co-Authored-By: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>