{
  "_provenance": "Hand-transcribed from the private analysis repo ~/rongorongo, runs 000-008, all committed. Each value is sourced to its run. This is the desk's own analysis, not an OLS fit to a public dataset; see the in-page method note.",
  "corpus": {
    "script_tokens": 11003,
    "script_base_signs": 14977,
    "tablet_codes": 24,
    "witness_texts_covered": "25/26",
    "witness_note": "lxgf Barthel DB (github.com/lxgf/rongorongo) cross-checked against independent CEIPP transliterations + Spaelti glyph-XML from kohaumotu.org; text Z (Poike) is illegible in the corpus itself.",
    "language_old_register_words": 12500,
    "language_old_register_note": "chants (~2,508) + Metraux 1940 old-register extraction (10,039, PD HathiTrust scan)",
    "language_mid_register_words": 18831,
    "language_mid_register_note": "Kieviet 2017 interlinear examples, CC-BY"
  },
  "system_type": {
    "jsd_sign_vs_syllable": 0.0160,
    "jsd_sign_vs_word": 0.0162,
    "jsd_syllable_vs_word": 0.0381,
    "coverage_90pct_syllables": 30,
    "coverage_90pct_signs": 150,
    "coverage_90pct_words": 352,
    "finding": "The frequency-curve argument that has anchored the syllabary case cannot tell syllabary from logography; the sign curve is equidistant from both. What discriminates is inventory demand: 150 signs to cover 90% of text sits between a syllabary's ~30 and logography's 352+. Mixed logosyllabic."
  },
  "effect": {
    "analyses_total": 172,
    "analyses_positive": 172,
    "analyses_note": "held-out split x language-model x aligner analyses across 3 independently built pipelines (run 000 chat, run 001 rebuild Phase A+B, run 005); observed alignment beat every run-preserving surrogate in every analysis",
    "representative_z": 8.0,
    "z_ranges_note": "run 001B Phase B: HC mean +8.84, EM +8.34; run 005: HC +7.83/+8.23 OLD/BLEND, EM +7.60/+7.94",
    "surrogates_per_split": 24
  },
  "values": {
    "candidates_at_start": 25,
    "validated": [
      {"sign": "001", "value": "a", "note": "modal in every split of every pipeline (five pipelines); won 40/40 in the frozen head-to-head vs Davletshin's published 001=ki"},
      {"sign": "064", "value": "na", "note": "18/20 EM splits under both language models; lurked sub-threshold (9/20) in every Phase B table before old-register power"}
    ],
    "revoked": {"sign": "003", "was": "ta", "walk": "e -> u -> ta -> i", "modal_now": "i", "modal_now_rate_em_old": "7/20", "modal_now_rate_em_blend": "9/20", "note": "met the Phase B stability bar; revoked at old-register power when its value would not hold"},
    "head_to_head": {
      "sign": "001", "ours": "a", "theirs": "ki", "theirs_source": "Davletshin 2022 JPS",
      "wins_old_lm": "20/20", "wins_blend_lm": "20/20",
      "margin_mean_old": 0.261, "margin_mean_blend": 0.299,
      "free_optimizer_chose_a": "20/20 under all four map tables; never ki",
      "note": "forcing 001=ki costs about as much held-out likelihood as the whole alignment effect gains over surrogates"
    },
    "unresolved": [{"sign": "381", "value": "ka", "note": "wins 5/20 OLD, 9/20 BLEND, near-zero margins"}],
    "disfavored": [
      {"sign": "010", "value": "a", "note": "0/20 both LMs"},
      {"sign": "002", "value": "e", "note": "0/20"},
      {"sign": "600", "value": "ma", "note": "0/20"}
    ]
  },
  "anchors": {
    "mamari": {"finding": "NULL", "p_shuffled_lexicon": 0.554, "p_shuffled_groups": 0.621, "perms": 10000, "crescents": 28, "note": "the lunar-calendar crescent groups do not spell the attested night names under any calendar rotation; supports Guy 1990's astronomical/pictorial reading"},
    "genealogy": {"finding": "one surviving claim after adversarial verification", "chain": "280 -> 730 -> 517 -> 222", "p_range": "0.003-0.026", "note": "the Butinov-Knorozov chain is corpus-unique, witness-robust, survives every null incl. repetition-preserving; the Santiago Staff (23% marker density) does not chain (p=0.23). The flashier supporting statistics were refuted: 076-density p=7.5e-9 collapsed to 0.055-0.164 under cluster-respecting nulls; periodicity fired only in the exact literature window."},
    "metoro_run004": {"finding": "NULL", "grammatical_glosses": "0/47", "fisher_p": 1.0, "note": "the 1873 informant Metoro rendered zero core signs as grammatical material; he narrativized everything"},
    "metoro_run008": {"finding": "NULL by non-recoverability", "note": "Metraux 1940 p.397 reading of tablet B row 21 (=Bv11, certain) could not be paired glyph-to-phrase at row resolution; Metraux himself calls the readings 'hasty explanations of the drawings'"}
  },
  "davletshin": {
    "convergence": "independently concludes mixed logosyllabic + East Polynesian by iconographic/parallel-text methods",
    "agreement": "his in-core syllabic signs are exactly our q~1.0 syllabographic block",
    "conflict_001": "his 001=ki vs our 001=a, resolved 40/40 for a on held-out distributional evidence (both frameworks agree 001 is syllabic)",
    "conceded": "sign 999 class shifted toward his logogram claim at old-register power (Phase B's 18/20 syllabographic no longer citable)",
    "source": "Davletshin 2022, Journal of the Polynesian Society 131(2):185-220, DOI 10.15286/jps.131.2.185-220"
  },
  "method_lesson": {
    "finding": "iid-sign nulls are anti-conservative by ~7 orders of magnitude in this corpus",
    "example": "the genealogy 076-density statistic read p=7.5e-9 under an iid binomial and p=0.055-0.164 once text-level sign clustering was respected",
    "rule": "only repetition/cluster-preserving nulls count; every positive claim carries one"
  }
}