{
  "schema_version": 1,
  "protocol": "Single-message public ONNX runtime; strict lexical accuracy excludes neutral and ambiguous tokens; exact cases accept allowed alternative labels.",
  "annotation_status": "AI-authored provisional rubric, user-supplied annotations and AI contextual review; not independent expert gold",
  "stress_suites": {
    "conversation_regression": {
      "suite_sha256": "de879b6b7c1c4c28a44bef7882da6396ddd2b11a08ec1db1d90435bf67b8eae3",
      "fp32": {
        "cases": 100,
        "strict_lexical_tokens": 640,
        "ambiguous_tokens": 13,
        "strict_lexical_accuracy": 0.9234375,
        "exact_case_accuracy": 0.73,
        "high_confidence_errors": 14,
        "per_class": {
          "O": {
            "precision": 0.9938837920489296,
            "recall": 1.0,
            "f1": 0.9969325153374233,
            "support": 325
          },
          "BA": {
            "precision": 0.9377162629757786,
            "recall": 0.9124579124579124,
            "f1": 0.924914675767918,
            "support": 297
          },
          "RU": {
            "precision": 0.9788732394366197,
            "recall": 0.9423728813559322,
            "f1": 0.9602763385146805,
            "support": 295
          },
          "TAT": {
            "precision": 0.5142857142857142,
            "recall": 0.75,
            "f1": 0.6101694915254237,
            "support": 24
          },
          "OTHER": {
            "precision": 0.8,
            "recall": 1.0,
            "f1": 0.888888888888889,
            "support": 24
          }
        }
      },
      "int8": {
        "cases": 100,
        "strict_lexical_tokens": 640,
        "ambiguous_tokens": 13,
        "strict_lexical_accuracy": 0.925,
        "exact_case_accuracy": 0.73,
        "high_confidence_errors": 15,
        "per_class": {
          "O": {
            "precision": 0.9938837920489296,
            "recall": 1.0,
            "f1": 0.9969325153374233,
            "support": 325
          },
          "BA": {
            "precision": 0.9319727891156463,
            "recall": 0.9225589225589226,
            "f1": 0.9272419627749577,
            "support": 297
          },
          "RU": {
            "precision": 0.9822064056939501,
            "recall": 0.9355932203389831,
            "f1": 0.9583333333333334,
            "support": 295
          },
          "TAT": {
            "precision": 0.5294117647058824,
            "recall": 0.75,
            "f1": 0.6206896551724139,
            "support": 24
          },
          "OTHER": {
            "precision": 0.8275862068965517,
            "recall": 1.0,
            "f1": 0.9056603773584906,
            "support": 24
          }
        }
      }
    },
    "borderline_conversations": {
      "suite_sha256": "c9c5439627b1e4f74fcf481b30c3f291e9d2f3a763bdd370a1491cd924094ce1",
      "fp32": {
        "cases": 40,
        "strict_lexical_tokens": 263,
        "ambiguous_tokens": 0,
        "strict_lexical_accuracy": 0.9239543726235742,
        "exact_case_accuracy": 0.675,
        "high_confidence_errors": 7,
        "per_class": {
          "O": {
            "precision": 1.0,
            "recall": 1.0,
            "f1": 1.0,
            "support": 115
          },
          "BA": {
            "precision": 0.9357798165137615,
            "recall": 0.9026548672566371,
            "f1": 0.918918918918919,
            "support": 113
          },
          "RU": {
            "precision": 0.95,
            "recall": 0.9743589743589743,
            "f1": 0.9620253164556962,
            "support": 117
          },
          "TAT": {
            "precision": 0.75,
            "recall": 0.8181818181818182,
            "f1": 0.7826086956521738,
            "support": 22
          },
          "OTHER": {
            "precision": 0.9,
            "recall": 0.8181818181818182,
            "f1": 0.8571428571428572,
            "support": 11
          }
        }
      },
      "int8": {
        "cases": 40,
        "strict_lexical_tokens": 263,
        "ambiguous_tokens": 0,
        "strict_lexical_accuracy": 0.9315589353612167,
        "exact_case_accuracy": 0.725,
        "high_confidence_errors": 9,
        "per_class": {
          "O": {
            "precision": 1.0,
            "recall": 1.0,
            "f1": 1.0,
            "support": 115
          },
          "BA": {
            "precision": 0.9217391304347826,
            "recall": 0.9380530973451328,
            "f1": 0.9298245614035088,
            "support": 113
          },
          "RU": {
            "precision": 0.9655172413793104,
            "recall": 0.9572649572649573,
            "f1": 0.9613733905579399,
            "support": 117
          },
          "TAT": {
            "precision": 0.8181818181818182,
            "recall": 0.8181818181818182,
            "f1": 0.8181818181818182,
            "support": 22
          },
          "OTHER": {
            "precision": 0.9,
            "recall": 0.8181818181818182,
            "f1": 0.8571428571428572,
            "support": 11
          }
        }
      }
    },
    "new_household_conversations": {
      "suite_sha256": "da216b8d4b411a595fe842b11421d6b793fedd38c85c89963c3adf72fcd3f282",
      "fp32": {
        "cases": 40,
        "strict_lexical_tokens": 344,
        "ambiguous_tokens": 0,
        "strict_lexical_accuracy": 0.936046511627907,
        "exact_case_accuracy": 0.75,
        "high_confidence_errors": 16,
        "per_class": {
          "O": {
            "precision": 1.0,
            "recall": 1.0,
            "f1": 1.0,
            "support": 104
          },
          "BA": {
            "precision": 0.925,
            "recall": 0.9098360655737705,
            "f1": 0.9173553719008265,
            "support": 122
          },
          "RU": {
            "precision": 0.9736842105263158,
            "recall": 0.9946236559139785,
            "f1": 0.9840425531914894,
            "support": 186
          },
          "TAT": {
            "precision": 0.8125,
            "recall": 0.7222222222222222,
            "f1": 0.7647058823529411,
            "support": 36
          },
          "OTHER": {
            "precision": 0.0,
            "recall": 0,
            "f1": 0,
            "support": 0
          }
        }
      },
      "int8": {
        "cases": 40,
        "strict_lexical_tokens": 344,
        "ambiguous_tokens": 0,
        "strict_lexical_accuracy": 0.9476744186046512,
        "exact_case_accuracy": 0.825,
        "high_confidence_errors": 16,
        "per_class": {
          "O": {
            "precision": 1.0,
            "recall": 1.0,
            "f1": 1.0,
            "support": 104
          },
          "BA": {
            "precision": 0.9333333333333333,
            "recall": 0.9180327868852459,
            "f1": 0.9256198347107439,
            "support": 122
          },
          "RU": {
            "precision": 0.9789473684210527,
            "recall": 1.0,
            "f1": 0.9893617021276596,
            "support": 186
          },
          "TAT": {
            "precision": 0.8235294117647058,
            "recall": 0.7777777777777778,
            "f1": 0.7999999999999999,
            "support": 36
          },
          "OTHER": {
            "precision": 0,
            "recall": 0,
            "f1": 0,
            "support": 0
          }
        }
      },
      "fp32_pytorch_parity": {
        "cases": 40,
        "max_probability_delta": 5.066394805908203e-06,
        "disagreement_tokens": 0,
        "passed": true
      }
    }
  },
  "selection_validation": {
    "earlier_annotations": {
      "rows": 139,
      "masked_tokens": 13,
      "lexical_accuracy": 0.9244,
      "used_for_epoch_selection": true
    },
    "borderline_annotations": {
      "rows": 77,
      "masked_tokens": 56,
      "lexical_accuracy": 0.9376000000000001,
      "used_for_epoch_selection": true
    },
    "new_household_annotations": {
      "rows": 61,
      "masked_tokens": 33,
      "lexical_accuracy": 0.9043000000000001,
      "used_for_epoch_selection": true
    }
  },
  "limitations": [
    "Authored stress tests do not measure natural traffic accuracy.",
    "Previously inspected suites are regression tests; new household suite has two context variants per independent core.",
    "No-diacritic BA/TAT text can be intrinsically ambiguous.",
    "Shared words, Russian-root inflections and short switches can be mislabeled with high confidence."
  ]
}