{
  "release": "audio-v3",
  "reference_release": "experimental-context-v2",
  "reference_weights_sha256": "8e0be4ecb637a35c4f152e2cb887bca8e95ab1f1b17daa4f62a1706331f7aa3e",
  "source_checkpoint_sha256": "f54918eb8146ee85fa429f1dffe30cf8c2adce3132b057fe2d9adc35b1f311ba",
  "stress_checkpoint_sha256": "5342342f6469de3f985dbc0799a3083f2ed8f979488f920812fb6a911125bd54",
  "stress_weights_bitwise_preserved": true,
  "input_yo_hidden_for_paired_comparison": true,
  "stress_checks": {
    "RHD_manual": {
      "label_kind": "manual contextual reference",
      "targets": 7676,
      "context_v2_correct": 7482,
      "audio_v3_correct": 7484,
      "context_v2_score": 0.9747264200104221,
      "audio_v3_score": 0.9749869723814487,
      "context_v2_macro": 0.9637496548122184,
      "audio_v3_macro": 0.9679068553070598
    },
    "external_russtress": {
      "label_kind": "public reference, previously reused during development",
      "targets": 11089,
      "context_v2_correct": 10435,
      "audio_v3_correct": 10483,
      "context_v2_score": 0.9410226350437371,
      "audio_v3_score": 0.9453512489854811,
      "context_v2_macro": 0.9361600330721678,
      "audio_v3_macro": 0.9425269193674681
    },
    "TeraTTS/stress_dataset_sft_proza": {
      "label_kind": "supplied labels",
      "targets": 21234,
      "context_v2_correct": 19900,
      "audio_v3_correct": 19913,
      "context_v2_score": 0.9371762268060657,
      "audio_v3_score": 0.9377884524818687,
      "context_v2_macro": 0.8681086799998965,
      "audio_v3_macro": 0.8710729274781607
    },
    "TeraTTS/open_accent": {
      "label_kind": "supplied labels",
      "targets": 46994,
      "context_v2_correct": 43577,
      "audio_v3_correct": 44045,
      "context_v2_score": 0.9272885900327701,
      "audio_v3_score": 0.9372473081670001,
      "context_v2_macro": 0.8973879348784138,
      "audio_v3_macro": 0.9089840690228318
    },
    "TeraTTS/stress_dataset_sft_poetry": {
      "label_kind": "supplied labels",
      "targets": 5882,
      "context_v2_correct": 4981,
      "audio_v3_correct": 4995,
      "context_v2_score": 0.846820809248555,
      "audio_v3_score": 0.8492009520571234,
      "context_v2_macro": 0.811748154183098,
      "audio_v3_macro": 0.8215229111497144
    },
    "supplied_dialogue": {
      "label_kind": "RUAccent pseudo-label agreement",
      "targets": 42965,
      "context_v2_correct": 42483,
      "audio_v3_correct": 42685,
      "context_v2_score": 0.9887815663912487,
      "audio_v3_score": 0.9934830676131735,
      "context_v2_macro": 0.9561116028270102,
      "audio_v3_macro": 0.9954060892784302
    },
    "agent_authored": {
      "label_kind": "authored development diagnostics; not representative",
      "targets": 12,
      "context_v2_correct": 12,
      "audio_v3_correct": 7,
      "context_v2_score": 1.0,
      "audio_v3_score": 0.5833333333333334,
      "context_v2_macro": 1.0,
      "audio_v3_macro": 0.5833333333333334
    },
    "audio/portal_ru": {
      "label_kind": "filtered acoustic teacher agreement",
      "targets": 58,
      "context_v2_correct": 58,
      "audio_v3_correct": 57,
      "context_v2_score": 1.0,
      "audio_v3_score": 0.9827586206896551,
      "context_v2_macro": 1.0,
      "audio_v3_macro": 0.9827586206896551
    }
  },
  "manual_context_macro": {
    "context_v2": 0.9637496548122184,
    "audio_v3": 0.9679068553070598
  },
  "acoustic_teacher_agreement": {
    "context_v2": 0.9921076879932695,
    "audio_v3": 0.9965546252153359
  },
  "yo": {
    "experimental": true,
    "enabled_by_default": false,
    "validation_selected_threshold": 0.8,
    "known_manual_targets": {
      "tp": 736,
      "fp": 34,
      "fn": 49,
      "tn": 2818,
      "precision": 0.9558441558441558,
      "recall": 0.9375796178343949,
      "f1": 0.9466237942122186,
      "accuracy": 0.9771789936761067
    },
    "authored_target_diagnostic": {
      "correct": 39,
      "total": 45,
      "positive_correct": 24,
      "positive_total": 27,
      "negative_correct": 15,
      "negative_total": 18
    },
    "authored_whole_text_diagnostic": {
      "sentences": 45,
      "whole_sentence_correct": 36,
      "tp": 28,
      "fp": 6,
      "fn": 6,
      "precision": 0.8235294117647058,
      "recall": 0.8235294117647058
    },
    "diagnostic_representative": false,
    "passed_default_restoration_gate": false
  },
  "data": {
    "train_contexts": 472192,
    "validation_contexts": 54434,
    "test_contexts": 67048,
    "usable_audio_texts": 49996,
    "filtered_acoustic_word_targets": 323918,
    "unknown_labels_masked": true,
    "source_texts_redistributed": false
  },
  "label_provenance": {
    "clarified_on": "2026-10-28",
    "supplied_dialogue": {
      "stress_label_generator": "RUAccent",
      "provenance_evidence": "User confirmed that the Russian dialogue data was stressed by RUAccent.",
      "metric_kind": "RUAccent pseudo-label agreement",
      "independent_human_reference": false,
      "interpretation": "An increase measures closer agreement with RUAccent, including its errors. Held-out contexts do not make automatic labels an independent correctness reference.",
      "numeric_scores_changed": false
    },
    "audio_ru_dialogs": {
      "metric_kind": "Filtered TeraAnaly acoustic-teacher agreement",
      "teacher_training_caveat": "TeraAnaly used RUAccent-derived training labels.",
      "independent_human_reference": false
    },
    "reference_comparisons": {
      "sources": [
        "RHD_manual",
        "external_russtress"
      ],
      "interpretation": "Report separately from pseudo-label agreement. Public references have been reused during development and do not establish a score on the private benchmark."
    }
  },
  "limitations": [
    "Dialogue labels were generated by RUAccent; agreement includes teacher errors.",
    "Acoustic scores measure agreement with TeraAnaly, itself trained with RUAccent-derived labels.",
    "Public reference tests were reused during development; historical model/data overlap is unknown.",
    "Authored diagnostics are development regression checks, not a representative human benchmark.",
    "Optional ё placement can invent or miss ё and can disagree with predicted stress.",
    "No private benchmark examples or expected answers were used; no claim of beating RUAccent."
  ]
}