F-04:
    title: Word lengths follow a rhythm inside lines that stops at line breaks
    statement: >
      Neighbouring words in a line have related lengths (Moran I 0.157 vs shuffle q99 0.081); cross-line share <3 %; no
      language, verse or human-gibberish control reproduces it; no producer we built predicts it.
    map: [A5, A13, dot3]
    claims: [c-295, c-305, c-320, c-328, c-329, c-335, c-336, c-360, c-409, c-414, c-380]
    status: standing
    load_bearing: true
    rests_on: [X-03]
    lines:
      - {axis: data, by: "3 transliterations", ref: results_s21.txt, agrees: true}
      - {axis: impl, by: "Daybreak independent run (exact reproduction)", claim: c-320, agrees: true}
      - {axis: impl, by: "roach r3c, own code and segmenter, held-out pages: within-line only", ref: runs/roach-r3/results_r3.txt, claim: c-414, agrees: true}
      - {axis: physical, by: "s45 ink-derived word gaps (stronger, 0.254)", ref: results_s45.txt, claim: c-360, agrees: true}
      - {axis: method, by: "s65 drawing-gap removal", ref: results_s65.txt, claim: c-409, agrees: true}
      - {axis: split, by: "roach r3 held-out pages", ref: runs/roach-r3/results_r3.txt, agrees: true}
      - {axis: order, by: "s79: Moran order-free under every transform", ref: results_s79.txt, claim: c-421, agrees: true}
      - {axis: external, by: "Voynich Ninja thread (quimqu, nablator 2025): global lag-1 word-length r +0.16, line-break contrast; per-line -0.07 reproduced and shown to be short-line bias (shuffle -0.153, excess +0.082)", ref: runs/explore/wordlen_perline_bias_20261018.txt, agrees: true, unverified: true}
      - {axis: impl, by: "roach r12 blind: within-line lag-1 length correlation positive on 12/15 (ZL) and 11/15 (IT) blocks, none across line breaks; Kalevala metrical verse 15/15, other natural texts <= 4/15", ref: runs/roach-r12/results_r12.txt, claim: c-456, agrees: true, note: "replicates the rhythm; not unique to the manuscript (metrical verse has it)"}
    controls:
      - {what: "64 language, verse and gibberish sidecars on real lines", ref: results_s17.txt, matched: true}
    points:
      - {h: H-01, rel: against, strength: strong, why: "Latin pools -0.03..+0.04, verse negative"}
      - {h: H-08, rel: against, strength: moderate, why: "gibberish 0.098, p 0.73 on the within-line form"}
      - {h: H-07, rel: against, strength: moderate, why: "s57 local-memory producer gives 0.02"}