{
  "schema_version": 1,
  "publication_date": "2026-09-18",
  "suggested_title": "Fast speech inference depends on the workload",
  "summary": "An isolated Resonance2 offshoot reduced transcript errors and measured faster inference. Its affect ensemble failed the frozen F1 release gate, so the original affect model remains the default.",
  "status": {
    "main_version_changed": false,
    "service_deployed": false,
    "affect_default": "original",
    "affect_candidate": "experimental",
    "asr_quality_acceptance": true
  },
  "units": {
    "rates": "Fractions unless a field explicitly ends in percent or percentage_points.",
    "latency": "Milliseconds unless explicitly seconds.",
    "speedup": "Own-control latency divided by candidate latency; values below1 are slower.",
    "confidence_interval": "Paired percentile bootstrap of declared recording/speaker groups; descriptive historical evidence."
  },
  "architecture": {
    "bounded_decisions": {
      "inspiration": "Shared input work and typed decisions inspired by Jev/System One.",
      "trained_scope": "31 vocal affect/delivery attributes.",
      "flow": [
        "Compile reusable question schema",
        "Encode audio once for the affect path",
        "Reuse bounded readouts across slots",
        "Return typed answers and eligible calibration metadata"
      ],
      "primitives": [
        {
          "id": "choice",
          "meaning": "Distribution over supplied options; requires an appropriate task definition."
        },
        {
          "id": "score",
          "meaning": "Distribution over rubric levels and its expected score."
        },
        {
          "id": "noul",
          "meaning": "Binary yes/no marginal for an eligible canonical attribute."
        }
      ],
      "calibration_boundary": "Eligible canonical binary outputs have fitted calibration. Choice/score distributions and free-text aliases are not demonstrated calibrated arbitrary-task reasoning.",
      "free_text_boundary": "Positive aliases route to existing attributes; negation, temporal queries and unseen tasks were not established.",
      "references": [
        "typesafe_system_one",
        "calibration"
      ]
    },
    "asr": {
      "flow": [
        "Separate full stock acoustic encoder",
        "TDT sequence decoder",
        "Decoded transcript"
      ],
      "stock_model": "NVIDIA Parakeet TDT0.6B v2",
      "quality_decoder": "Beam8, native temperature1",
      "speed_decoder": "Greedy decoding",
      "weights_changed": false,
      "shared_with_affect_encoder": false,
      "reason": "The archives have different learned acoustic weights. No shared-tower or combined-endpoint result is claimed.",
      "references": [
        "parakeet_model",
        "exact_greedy_graphs"
      ]
    }
  },
  "cohorts": {
    "english_cv_development_200": {
      "recordings": 200,
      "speakers": 200,
      "reference_words": 1864,
      "audio_seconds": 1233.888,
      "purpose": "Decoder selection",
      "source": "English Common Voice development material",
      "historically_exposed": true
    },
    "english_cv_confirmation_500": {
      "recordings": 500,
      "speakers": 500,
      "reference_words": 4423,
      "audio_seconds": 2682.3656875,
      "purpose": "Frozen decoder confirmation and separately labelled latency follow-ups",
      "source": "English Common Voice development material",
      "historically_exposed": true,
      "fresh_official_test": false,
      "development_disjoint_by": [
        "speaker",
        "recording ID",
        "audio file hash",
        "decoded audio hash"
      ],
      "reference_note": "Human reading prompts previously normalized for CTC; no fresh independent adjudication."
    },
    "english_fleurs_20": {
      "recordings": 20,
      "reference_words": 382,
      "audio_seconds": 173.38,
      "purpose": "Separate development domain diagnostic and warm exact-shape latency fixture",
      "source": "English FLEURS",
      "fresh_test": false
    },
    "affect_historical_test": {
      "recordings": 7381,
      "recording_groups": 362,
      "attributes": 31,
      "conversation_recordings": 747,
      "conversation_groups": 145,
      "source": "Existing multi-corpus human-labelled research split",
      "historically_exposed": true,
      "fresh_unseen_test": false
    },
    "affect_splits": {
      "training_recordings": 19672,
      "validation_recordings": 4300,
      "calibration_recordings": 2945,
      "test_recordings": 7381,
      "pairwise_recording_id_overlap": 0,
      "pairwise_reference_group_overlap": 0
    }
  },
  "asr": {
    "profiles": {
      "reference": {
        "label": "Original control",
        "description": "Original-style stock greedy transcription.",
        "model_weights_changed": false
      },
      "quality": {
        "label": "Accuracy profile",
        "description": "Direct tensor inference, native-temperature beam8 TDT decoding and bounded encoder/decoder graphs.",
        "model_weights_changed": false
      },
      "speed": {
        "label": "Fast greedy profile",
        "description": "Direct tensor inference with greedy decoding and bounded encoder/decoder graphs.",
        "model_weights_changed": false
      }
    },
    "confirmation": {
      "reference": {
        "errors": 278,
        "reference_words": 4423,
        "substitutions": 227,
        "deletions": 27,
        "insertions": 24,
        "wer": 0.06285326701333936,
        "wer_percent": 6.285326701333936
      },
      "quality": {
        "errors": 264,
        "reference_words": 4423,
        "substitutions": 217,
        "deletions": 27,
        "insertions": 20,
        "wer": 0.059687994573818674,
        "wer_percent": 5.968799457381867,
        "delta_wer": -0.0031652724395206874,
        "delta_percentage_points": -0.31652724395206877,
        "relative_error_reduction_percent": 5.0359712230215825,
        "delta_wer_ci95": [
          -0.006262131477751327,
          -0.0004489136642396735
        ],
        "delta_percentage_points_ci95": [
          -0.6262131477751327,
          -0.04489136642396735
        ]
      },
      "speed": {
        "errors": 278,
        "reference_words": 4423,
        "substitutions": 227,
        "deletions": 27,
        "insertions": 24,
        "wer": 0.06285326701333936,
        "wer_percent": 6.285326701333936,
        "raw_transcript_matches": 500,
        "delta_wer": 0.0,
        "delta_wer_ci95": [
          0.0,
          0.0
        ]
      }
    },
    "acceptance_rule": "Quality WER must not exceed the matched original greedy control, with decoder identity frozen before confirmation scoring.",
    "quality_accepted": true,
    "uncertainty_note": "2000 paired speaker bootstrap draws; historical exposure and development selection remain. A nonzero observed improvement is not a generalization guarantee.",
    "normalization": "NFKC, casefold, Unicode punctuation/symbols to spaces, collapse whitespace; no number expansion.",
    "development": {
      "stock-greedy-encoder": {
        "errors": 128,
        "reference_words": 1864,
        "wer": 0.06866952789699571
      },
      "stock-beam8-native-encoder": {
        "errors": 124,
        "reference_words": 1864,
        "wer": 0.06652360515021459
      },
      "stock-beam4-native-encoder": {
        "errors": 132,
        "reference_words": 1864,
        "wer": 0.07081545064377683
      }
    },
    "latency_workloads": [
      {
        "id": "warm_exact_shapes",
        "label": "Warm exact shapes",
        "cohort": "english_fleurs_20",
        "recordings": 20,
        "repeats_per_recording": 10,
        "requests_per_profile": 200,
        "warmup_per_recording": 2,
        "scope": "Preloaded waveform to decoded text, CUDA synchronized, batch1; each exact shape already warmed.",
        "excludes": [
          "model loading",
          "file reads and decoding",
          "network and HTTP",
          "prewarming"
        ],
        "environment": "Same optional CUDA dependencies for control and candidates.",
        "profiles": {
          "reference": {
            "mean_ms": 73.43815858000426,
            "p50_ms": 72.3303570000553,
            "p95_ms": 82.56399265001163,
            "decoder_graph_calls": 200,
            "mean_speedup_vs_own_control": 1.0,
            "p95_speedup_vs_own_control": 1.0
          },
          "quality": {
            "mean_ms": 37.841653004988984,
            "p50_ms": 35.07269699991866,
            "p95_ms": 59.70063259982206,
            "decoder_graph_calls": 200,
            "encoder_replays": 200,
            "encoder_eager_fallbacks": 0,
            "cached_encoder_shapes": 19,
            "mean_speedup_vs_own_control": 1.9406699429943581,
            "p95_speedup_vs_own_control": 1.3829667970764135,
            "encoder_captures_during_measurement": 0
          },
          "speed": {
            "mean_ms": 18.36833865001381,
            "p50_ms": 17.58855500008849,
            "p95_ms": 25.970204199916143,
            "decoder_graph_calls": 200,
            "encoder_replays": 200,
            "encoder_eager_fallbacks": 0,
            "cached_encoder_shapes": 19,
            "mean_speedup_vs_own_control": 3.9980838756992885,
            "p95_speedup_vs_own_control": 3.179181496396483,
            "encoder_captures_during_measurement": 0
          }
        },
        "warning": "These gains require warm exact input shapes; this is a separate20-recording latency workload from the500-recording WER cohort.",
        "evidence": [
          "asr_summary",
          "asr_independent_audit"
        ]
      },
      {
        "id": "cold_mixed_shapes",
        "label": "No prewarm: mixed shapes",
        "cohort": "english_cv_confirmation_500",
        "recordings": 500,
        "repeats_per_recording": 1,
        "requests_per_profile": 500,
        "scope": "One ordered batch1 pass; waveform to decoded text with decoder setup and encoder graph captures included.",
        "excludes": [
          "model loading",
          "file reads and decoding",
          "network and HTTP"
        ],
        "environment": "Original reference dependency environment versus candidates with optional CUDA graph dependencies. Do not describe these environments as identical.",
        "profiles": {
          "reference": {
            "mean_ms": 72.22371625600952,
            "p95_ms": 82.17690665014742,
            "rtf": 0.013462690153057202,
            "total_inference_seconds": 36.11185812800476,
            "mean_speedup_vs_own_control": 1.0,
            "p95_speedup_vs_own_control": 1.0
          },
          "quality": {
            "mean_ms": 61.259963353991225,
            "p95_ms": 205.13582775026862,
            "rtf": 0.011419017854177503,
            "total_inference_seconds": 30.629981676995612,
            "encoder_replays": 239,
            "encoder_eager_fallbacks": 261,
            "mean_speedup_vs_own_control": 1.1789709347141486,
            "p95_speedup_vs_own_control": 0.40059753360192735,
            "encoder_captures_during_measurement": 32
          },
          "speed": {
            "mean_ms": 48.84525207799834,
            "p95_ms": 191.83087130011234,
            "rtf": 0.009104883108522529,
            "total_inference_seconds": 24.42262603899917,
            "encoder_replays": 239,
            "encoder_eager_fallbacks": 261,
            "mean_speedup_vs_own_control": 1.478623063315947,
            "p95_speedup_vs_own_control": 0.4283820747578458,
            "encoder_captures_during_measurement": 32
          }
        },
        "warning": "Single-pass diagnostic, not repeated latency estimation. Candidate cold p95 regresses.",
        "evidence": [
          "asr_summary",
          "asr_independent_audit"
        ]
      },
      {
        "id": "development_warmed_frozen_cache",
        "label": "Development-warmed, frozen cache",
        "cohort": "english_cv_confirmation_500",
        "recordings": 500,
        "repeats_per_recording": 3,
        "requests_per_profile": 1500,
        "scope": "Preloaded waveform to decoded text, batch1; same shuffled orders for all profiles, with development-only prewarming then a frozen32-shape encoder cache.",
        "excludes": [
          "model loading",
          "file reads and decoding",
          "network and HTTP",
          "development prewarming"
        ],
        "environment": "Same dependency environment for all three profiles.",
        "warm_recordings": 69,
        "warm_passes": 2,
        "cached_encoder_shapes": 32,
        "profiles": {
          "quality": {
            "mean_ms": 60.857166336664406,
            "p50_ms": 62.17078850022517,
            "p95_ms": 74.64511930024855,
            "rtf": 0.011343935433610485,
            "total_inference_seconds": 91.28574950499662,
            "per_pass": [
              {
                "samples": 500,
                "mean_ms": 63.74330848798854,
                "p95_ms": 79.72817274985573
              },
              {
                "samples": 500,
                "mean_ms": 59.80223450200174,
                "p95_ms": 71.53054939990396
              },
              {
                "samples": 500,
                "mean_ms": 59.02595602000292,
                "p95_ms": 71.1961848500323
              }
            ],
            "prior_raw_transcript_matches": 1500,
            "decoder_graph_calls": 1500,
            "decoder_reinitializations": 0,
            "encoder_replays": 159,
            "encoder_eager_fallbacks": 1341,
            "mean_speedup_vs_own_control": 1.1362393810909908,
            "p95_speedup_vs_own_control": 1.038601940448644,
            "encoder_captures_during_measurement": 0
          },
          "reference": {
            "mean_ms": 69.14830901332304,
            "p50_ms": 68.41886650045126,
            "p95_ms": 77.52656575025867,
            "rtf": 0.01288942617622174,
            "total_inference_seconds": 103.72246351998457,
            "per_pass": [
              {
                "samples": 500,
                "mean_ms": 71.93865194399041,
                "p95_ms": 82.17210585016801
              },
              {
                "samples": 500,
                "mean_ms": 68.64034930798334,
                "p95_ms": 72.12204409925107
              },
              {
                "samples": 500,
                "mean_ms": 66.86592578799537,
                "p95_ms": 70.19835754967971
              }
            ],
            "prior_raw_transcript_matches": 1500,
            "decoder_graph_calls": 1500,
            "decoder_reinitializations": 0,
            "mean_speedup_vs_own_control": 1.0,
            "p95_speedup_vs_own_control": 1.0
          },
          "speed": {
            "mean_ms": 48.48756823399102,
            "p50_ms": 50.818424000226514,
            "p95_ms": 59.38408664965209,
            "rtf": 0.009038209901794202,
            "total_inference_seconds": 72.73135235098654,
            "per_pass": [
              {
                "samples": 500,
                "mean_ms": 50.411951901993234,
                "p95_ms": 63.31095899990942
              },
              {
                "samples": 500,
                "mean_ms": 47.57655608001005,
                "p95_ms": 54.02707724974789
              },
              {
                "samples": 500,
                "mean_ms": 47.47419671996977,
                "p95_ms": 53.97808200027612
              }
            ],
            "prior_raw_transcript_matches": 1500,
            "decoder_graph_calls": 1500,
            "decoder_reinitializations": 0,
            "encoder_replays": 159,
            "encoder_eager_fallbacks": 1341,
            "mean_speedup_vs_own_control": 1.4261038763508935,
            "p95_speedup_vs_own_control": 1.3055107878924137,
            "encoder_captures_during_measurement": 0
          }
        },
        "warning": "Most shapes miss the cache and run eagerly. Quality pooled p95 improves slightly, but its third-pass p95 is worse than the control. Sequential profiles and per-pass drift limit timing inference.",
        "evidence": [
          "asr_frozen_cache_audit"
        ]
      }
    ],
    "hardware": "NVIDIA A100-SXM4-40GB",
    "batch_size": 1,
    "warm_vs_wer_cohort_note": "The20-clip warm latency fixture and500-clip WER confirmation are different cohorts; do not plot their accuracy/latency as one paired per-clip experiment.",
    "evidence": [
      "asr_summary",
      "asr_independent_audit",
      "asr_frozen_cache_audit"
    ]
  },
  "affect": {
    "profiles": {
      "original": {
        "label": "Original retained affect model",
        "status": "default",
        "encoder_precision": "FP32"
      },
      "experimental_ensemble": {
        "label": "Full-depth ensemble",
        "status": "experimental; release rejected",
        "encoder_precision": "BF16",
        "head_precision": "FP32",
        "mixture": "65% retained head and35% BCE replay head; both use one shared affect encoder"
      }
    },
    "historical_confirmation": {
      "original": {
        "all": {
          "recordings": 7381,
          "recording_groups": 362,
          "native_macro_f1": 0.41451817414150843,
          "native_macro_f1_percent": 41.451817414150845,
          "raw_macro_ap": 0.4346357830235026,
          "raw_macro_ap_percent": 43.46357830235026,
          "fraction_brier": 0.047476454618705824
        },
        "conversational": {
          "recordings": 747,
          "recording_groups": 145,
          "native_macro_f1": 0.17393567893186368,
          "native_macro_f1_percent": 17.39356789318637,
          "raw_macro_ap": 0.24919724025995862,
          "raw_macro_ap_percent": 24.919724025995862,
          "fraction_brier": 0.056678150258127405
        }
      },
      "experimental_ensemble": {
        "all": {
          "recordings": 7381,
          "recording_groups": 362,
          "native_macro_f1": 0.41422056471235696,
          "native_macro_f1_percent": 41.4220564712357,
          "raw_macro_ap": 0.43673286993370586,
          "raw_macro_ap_percent": 43.67328699337059,
          "fraction_brier": 0.04730176602026355
        },
        "conversational": {
          "recordings": 747,
          "recording_groups": 145,
          "native_macro_f1": 0.17295049537078216,
          "native_macro_f1_percent": 17.295049537078217,
          "raw_macro_ap": 0.24998565254926375,
          "raw_macro_ap_percent": 24.998565254926376,
          "fraction_brier": 0.056544205484715865
        }
      }
    },
    "metric_definitions": {
      "native_macro_f1": "Unweighted mean over all31 labels, including zero-support labels as zero; binary target is annotation fraction>=0.5; each model uses frozen native thresholds.",
      "raw_macro_ap": "Mean average precision over attributes having both observed classes; ranking metric.",
      "fraction_brier": "Squared error against annotation fractions, averaging observed labels within each clip then observed clips. Lower is better."
    },
    "uncertainty": {
      "all": {
        "draws": 2000,
        "recording_groups": 362,
        "native_macro_f1_delta": -0.0002976094291514686,
        "native_macro_f1_delta_ci95": [
          -0.0019391602620193473,
          0.002031689767151009
        ],
        "native_macro_f1_delta_percentage_points": -0.02976094291514686,
        "native_macro_f1_delta_percentage_points_ci95": [
          -0.19391602620193474,
          0.20316897671510092
        ],
        "fraction_brier_delta": -0.00017468859844227413,
        "fraction_brier_delta_ci95": [
          -0.00022032685220067636,
          -0.00013788531046367086
        ]
      },
      "conversational": {
        "draws": 2000,
        "recording_groups": 145,
        "native_macro_f1_delta": -0.0009851835610815152,
        "native_macro_f1_delta_ci95": [
          -0.005674063081023163,
          0.0038898445843030334
        ],
        "native_macro_f1_delta_percentage_points": -0.09851835610815152,
        "native_macro_f1_delta_percentage_points_ci95": [
          -0.5674063081023163,
          0.3889844584303033
        ],
        "fraction_brier_delta": -0.0001339447734115401,
        "fraction_brier_delta_ci95": [
          -0.00020905740247190128,
          -6.635799089070497e-05
        ]
      }
    },
    "release_accepted": false,
    "release_checks": {
      "all31_calibrated_fraction_brier_maximum_delta": {
        "bound": 0.0002,
        "delta": -0.00017468859844227413,
        "passed": true
      },
      "all31_calibrated_native_macro_f1_minimum_delta": {
        "bound": 0.0,
        "delta": -0.0002976094291514686,
        "passed": false
      },
      "all31_raw_macro_ap_minimum_delta": {
        "bound": -0.001,
        "delta": 0.002097086910203272,
        "passed": true
      },
      "conversation_calibrated_native_macro_f1_minimum_delta": {
        "bound": -0.001,
        "delta": -0.0009851835610815152,
        "passed": true
      }
    },
    "release_decision": "Original remains default; candidate stays experimental. No weights, precision or thresholds were reselected after confirmation.",
    "precision_development_accepted": true,
    "precision_development_checks": {
      "all31_calibrated_fraction_brier_maximum_delta": {
        "bound": 0.0002,
        "delta": 1.224468003992868e-05,
        "passed": true
      },
      "all31_calibrated_native_macro_f1_minimum_delta": {
        "bound": -0.001,
        "delta": -0.000490871407323179,
        "passed": true
      },
      "all31_raw_macro_ap_minimum_delta": {
        "bound": -0.001,
        "delta": -6.791947264028764e-05,
        "passed": true
      },
      "conversation_calibrated_native_macro_f1_minimum_delta": {
        "bound": -0.002,
        "delta": 0.00016537566355454003,
        "passed": true
      }
    },
    "warm_latency": {
      "hardware": "NVIDIA A100-SXM4-40GB",
      "batch_size": 1,
      "canonical_queries": 31,
      "recordings": 1,
      "iterations_per_duration": 100,
      "warmup_per_duration": 10,
      "scope": "Real waveform prefixes from one recording to CPU canonical probabilities; warmed. This measures compute scaling, not varied production traffic.",
      "excludes": [
        "recording time",
        "file decoding and resampling",
        "model loading",
        "text/schema compilation",
        "network and HTTP",
        "queues and JSON formatting"
      ],
      "cases": [
        {
          "duration_seconds": 5,
          "original": {
            "mean_ms": 18.46155138,
            "p50_ms": 18.372774999999997,
            "p95_ms": 19.1373601
          },
          "experimental_bf16": {
            "mean_ms": 14.565230199999998,
            "p50_ms": 14.5244725,
            "p95_ms": 14.99759345
          },
          "same_ensemble_fp32": {
            "mean_ms": 19.27670005,
            "p50_ms": 19.1460195,
            "p95_ms": 20.068240799999998
          }
        },
        {
          "duration_seconds": 10,
          "original": {
            "mean_ms": 23.99546666,
            "p50_ms": 23.95091,
            "p95_ms": 24.614379
          },
          "experimental_bf16": {
            "mean_ms": 15.35548452,
            "p50_ms": 15.3317135,
            "p95_ms": 15.7564102
          },
          "same_ensemble_fp32": {
            "mean_ms": 25.134081410000004,
            "p50_ms": 25.030258500000002,
            "p95_ms": 25.8883516
          }
        },
        {
          "duration_seconds": 20,
          "original": {
            "mean_ms": 36.539131860000005,
            "p50_ms": 36.620948999999996,
            "p95_ms": 36.98808245
          },
          "experimental_bf16": {
            "mean_ms": 20.01013626,
            "p50_ms": 19.963167,
            "p95_ms": 20.3665954
          },
          "same_ensemble_fp32": {
            "mean_ms": 37.51481867,
            "p50_ms": 37.508576,
            "p95_ms": 37.6762042
          }
        }
      ],
      "evidence": [
        "affect_bf16_latency",
        "affect_fp32_latency"
      ]
    },
    "evidence": [
      "affect_release_gate",
      "affect_uncertainty",
      "affect_independent_audit"
    ],
    "development_waveform": {
      "original_fp32": {
        "all": {
          "recordings": 4300,
          "recording_groups": 253,
          "native_macro_f1": 0.579731553091457,
          "native_macro_f1_percent": 57.973155309145696,
          "raw_macro_ap": 0.6065533264460113,
          "fraction_brier": 0.03830227395689518
        },
        "conversational": {
          "recordings": 746,
          "recording_groups": 168,
          "native_macro_f1": 0.18058943250541395,
          "native_macro_f1_percent": 18.058943250541397,
          "raw_macro_ap": 0.2504592754072191,
          "fraction_brier": 0.057381121362142476
        }
      },
      "ensemble_fp32": {
        "all": {
          "recordings": 4300,
          "recording_groups": 253,
          "native_macro_f1": 0.5826520534315003,
          "native_macro_f1_percent": 58.26520534315003,
          "raw_macro_ap": 0.6083940149636107,
          "fraction_brier": 0.038125363217348614
        },
        "conversational": {
          "recordings": 746,
          "recording_groups": 168,
          "native_macro_f1": 0.18081158183233023,
          "native_macro_f1_percent": 18.081158183233022,
          "raw_macro_ap": 0.2504280107534496,
          "fraction_brier": 0.05728416245347328
        }
      },
      "ensemble_bf16": {
        "all": {
          "recordings": 4300,
          "recording_groups": 253,
          "native_macro_f1": 0.5821611820241771,
          "native_macro_f1_percent": 58.21611820241771,
          "raw_macro_ap": 0.6083260954909704,
          "fraction_brier": 0.03813760789738854
        },
        "conversational": {
          "recordings": 746,
          "recording_groups": 168,
          "native_macro_f1": 0.18097695749588477,
          "native_macro_f1_percent": 18.09769574958848,
          "raw_macro_ap": 0.25090138999374606,
          "fraction_brier": 0.05729966583659739
        }
      }
    }
  },
  "interaction_contract": {
    "default_task": "asr",
    "default_latency_workload": "development_warmed_frozen_cache",
    "profile_ids": [
      "reference",
      "quality",
      "speed"
    ],
    "latency_metric_options": [
      "mean_ms",
      "p95_ms"
    ],
    "show_evidence_cohort_next_to_every_chart": true,
    "show_failed_affect_gate_inline": true,
    "display_intervals_in_percentage_points": true,
    "do_not_interpolate_unmeasured_durations": true,
    "do_not_combine_affect_and_asr_latency": true,
    "do_not_publish_raw_predictions_or_audio": true,
    "do_not_use_unlabelled_confidence_badges": true,
    "warm_and_cold_workloads_must_remain_separate": true,
    "controls_update_claim_and_caveat_together": true
  },
  "limitations": [
    "English, short whole-utterance recordings; no streaming, multilingual or long-form result.",
    "Historical cohorts, no fresh unseen-test claim.",
    "Latency is GPU inference with preloaded waveform input; no network or hosted-endpoint benchmark.",
    "The Jev-inspired bounded decision interface and the TDT transcript decoder are separate mechanisms.",
    "Lower WER does not imply higher affect F1. The affect candidate failed its release gate.",
    "Reported graph speedups depend on workload and cache policy; measured eager misses remain part of runtime."
  ],
  "references": [
    {
      "id": "typesafe_system_one",
      "title": "System One — TypeSafe documentation",
      "url": "https://docs.typesafe.ai/concepts/system-one",
      "supports": "Typed choice, score and noul decisions; Jev input is text only. Interface inspiration, not evidence about this speech system.",
      "verified_on": "2026-09-18"
    },
    {
      "id": "parakeet_model",
      "title": "NVIDIA Parakeet TDT 0.6B v2 model card",
      "url": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2",
      "supports": "Stock English ASR model combines a FastConformer encoder with a TDT decoder. Our results are separate from published model-card benchmarks.",
      "verified_on": "2026-09-18"
    },
    {
      "id": "exact_greedy_graphs",
      "title": "Speed of Light Exact Greedy Decoding for RNN-T Speech Recognition Models on GPU",
      "url": "https://arxiv.org/abs/2406.03791",
      "supports": "Prior work on efficient exact greedy GPU decoding; no paper speedup is claimed as our measurement.",
      "verified_on": "2026-09-18"
    },
    {
      "id": "calibration",
      "title": "On Calibration of Modern Neural Networks",
      "url": "https://proceedings.mlr.press/v70/guo17a.html",
      "supports": "Post-training probability calibration is a separate empirical step, not a property guaranteed by a softmax output.",
      "verified_on": "2026-09-18"
    }
  ],
  "evidence_fingerprints": {
    "asr_summary": {
      "kind": "measured project artifact",
      "sha256": "995a9d19b23d4f3297215782ee4b544b59a4411af49ad919b51ce76344e81274"
    },
    "asr_independent_audit": {
      "kind": "measured project artifact",
      "sha256": "629d467011c767667ddce6a1a80768bec03b6951f3b910563da9ff3c4185de83"
    },
    "asr_frozen_cache_audit": {
      "kind": "measured project artifact",
      "sha256": "bf17caf9f2f5bfec261296cda60405fa4aa92d18d45ff43d2d372e39af7e752d"
    },
    "affect_release_gate": {
      "kind": "measured project artifact",
      "sha256": "5de6981a9f8df47d266e23f5c121b2eb1cd9710a3fa491f52ff3a31dc2ea7f73"
    },
    "affect_uncertainty": {
      "kind": "measured project artifact",
      "sha256": "e123848aaf1af8a8531fe0b2c901d7a842c032279a0665212ccbe8e7c77a0e83"
    },
    "affect_precision_gate": {
      "kind": "measured project artifact",
      "sha256": "6f54331c6266179545768f67afeebae30f9e51d2107050dbb36272a7df39992c"
    },
    "affect_independent_audit": {
      "kind": "measured project artifact",
      "sha256": "c1b9e9caa69c25eed599c7e627cb9125541d87c28bf194aee5f24894f7b81d09"
    },
    "affect_bf16_latency": {
      "kind": "measured project artifact",
      "sha256": "34d6ba4c5cdeaed746129743d29e4dd849bb42517a3e6bde2f3c7d6038947947"
    },
    "affect_fp32_latency": {
      "kind": "measured project artifact",
      "sha256": "047d28b8b4b5e019899a459040ad80cf42206df96ee2e1711c9e1059f1ac7dfd"
    },
    "asr_development": {
      "kind": "measured project artifact",
      "sha256": "20ee9b97ffb2a6e7248ba1ce745f42a63454eb64396114f9dc0a4270865c6d6d"
    },
    "asr_warm_reference": {
      "kind": "measured project artifact",
      "sha256": "f7676a81ded5f20079386ca69b16d567fed90b94f6b3e69c0b73fbee2cf495e6"
    },
    "asr_warm_quality": {
      "kind": "measured project artifact",
      "sha256": "740e71e97322f4f33bf2cbcf752020988ab51ae6d35113b865e9c20b04880fbc"
    },
    "asr_warm_speed": {
      "kind": "measured project artifact",
      "sha256": "9e477b05f7c4e04d273d353a1dea618bf7c09ab65f225378efa27e581e2453ed"
    }
  },
  "public_safety": {
    "content": "Aggregate metrics only",
    "omitted": [
      "raw transcripts",
      "recording and speaker identifiers",
      "audio hashes and file names",
      "absolute local paths",
      "remote paths",
      "cloud identifiers",
      "commands and credentials",
      "raw prediction archives"
    ],
    "experiment_repository_modified": false
  }
}