{
  "config": {
    "schema": "q1-v2-source-first-2026-09",
    "source_cache_schema": "q1-b0b4-v4-openface",
    "common_step_s": 0.1,
    "multi_context_windows_s": [
      0.1,
      0.3,
      0.7
    ],
    "relative_progress_bins": 50,
    "text_encoder": "google-bert/bert-base-uncased; frozen last-four-layer mean per whitespace word; window=510, stride=384",
    "alignment": "fixed-transcript CTC monotone state graph; hard Viterbi path plus stored forward-backward occupancy; scores uncalibrated",
    "audio": "74-D: 40 log-Mel + 13 MFCC + 13 two-sided local-linear delta MFCC + [log-energy, log-F0, librosa.pyin voiced probability, spectral centroid, bandwidth, flux, zero-crossing rate, HNR dB]; 16 kHz, 400-sample window, 160-sample hop, FFT 512; HNR autocorrelation uses n=1024 at detected F0; preserve per-field validity; librosa.pyin warns cycle support is short at fmin=80 Hz for a 25 ms frame",
    "vision": "OpenFace 2.2.0 native-frame output: 17 AU intensities, Pose6, Gaze6, Geometry6; confidence threshold 0.8; actual frame PTS and indices retained",
    "quality": "B0 uses raw CTC word-path score in [0,1]; unavailable quality falls back to q*=1 and is flagged; audio/vision retain explicit unknown or detector-confidence states",
    "content_probe": "disabled; physical time/query maps remain separate from content similarity",
    "time": "left-closed/right-open seconds relative to video stream origin; source indices retained",
    "branch_states": {
      "sec": "primary B0 view and materialized",
      "B1_vision_geometry": "geometry-aware 0.1 s view materialized",
      "word": "materialized",
      "phase50": "materialized",
      "multi": "derived from native rows on demand by q1_io",
      "posterior": "native CTC posterior stored; sec projection available through q1_io",
      "dynamics": "materialized if continuous support is sufficient",
      "b4_speech": "source-cache-backed on demand through q1_io to stay within the 50 MiB feature-package cap",
      "query_word": "physical H_time matrices materialized",
      "content_probe": "disabled"
    }
  },
  "config_hash": "c1b931d6d64931209313a8b2f0e501069e2056939a15e900c1ee0762ba3c0617",
  "source_run_manifest": "source_cache_manifest.json",
  "source_cache_schema": "q1-b0b4-v4-openface",
  "source_label_sha256": "827334f782b1f242c84ad944a657bc7f7c63643f71ae9c977811a3eecdd3ab33",
  "python": "3.14.7 (main, Aug 10 2026, 00:00:00) [GCC 16.1.1 20260515 (Red Hat 16.1.1-2)]",
  "packages": {
    "torch": "2.14.0+cu130",
    "transformers": "5.17.0",
    "numpy": "2.5.3",
    "scipy": "1.18.1",
    "scikit-learn": "1.9.1",
    "opencv-python-headless": "5.0.0.93",
    "librosa": "1.0.0",
    "openpyxl": "3.1.5",
    "matplotlib": "3.11.2"
  },
  "models": {
    "text_id": "google-bert/bert-base-uncased",
    "text_revision": "86b5e0934494bd15c9632b12f734a8a67f723594",
    "speech_id": "facebook/wav2vec2-base-960h",
    "speech_revision": "22aad52d435eb6dbaf354bdad9b0da84ce7d6156",
    "speech_stride_samples": 320,
    "speech_receptive_field_samples": 400
  },
  "samples": 100,
  "modalities": 300,
  "total_bytes": 44321567
}