{
  "cases": 10,
  "languages": {
    "en": "English",
    "de": "German"
  },
  "models": {
    "base17": "Humaneness Voice Large 1.7",
    "podcast": "Humaneness Podcast full fine-tune · step 9,874"
  },
  "raw_candidates": 1536,
  "episode_conditions": 160,
  "reference": "de-notebooklm-new",
  "reference_language": "German for both English and German outputs",
  "strategies": {
    "segment": "One active-speaker reference; a fresh single-speaker generation per turn, identical native format on both models",
    "block": "Four or five turns at a time, both fixed identity anchors, no past dialogue or generated-audio feedback; baseline1.7multi-speaker format is exploratory"
  },
  "timing": [
    "Natural-rate estimated duration (about180English/168Germanwords per minute)",
    "No inline target duration and global Tokens=None"
  ],
  "cfg": 2.5,
  "temperature": 1.0,
  "top_p": 0.95,
  "top_k": 50,
  "segment_candidate_count": 1,
  "block_candidate_count": 8,
  "block_generation_batch": 4,
  "nested_best_of": [
    1,
    4,
    8
  ],
  "gap_candidates_ms": [
    100,
    200,
    300,
    400,
    500
  ],
  "block_reward": "R = (0.5*C +0.5*M)/(1+WER). C is speaker-balanced mean own-reference identity cosine clipped to[0,1]. M is speaker-balanced positive own-versus-other reference margin, normalized by max(1-reference_cross_cosine,.05), clipped to[0,1]. Unaligned turns contribute zero. No speaker relabelling to hide swaps.",
  "gap_reward": "Score each complete assembled waveform. Report G-only, CE-only and50/50average of their fixed EARS training-normalized Gaussian CDFs. Always retain fixed400ms.",
  "gap_clock": "Desired gap is last ASR word end to next first ASR word start. Preserve30ms guards and nearby confident nonverbal events; subtract retained guards from inserted silence. Skip infeasible gaps and label any minimum-feasible rescue.",
  "assembly": "Keep internal block timing unchanged; only newly created unit boundaries are gap-optimized. Five-ms edge fades stay outside detected speech/bursts.",
  "references_included_in_input": "Only current identity anchor for segment mode; both identity anchors for block mode. No context montage.",
  "duration_caps_seconds": {
    "segment": 30,
    "block": 60
  },
  "candidate_retention": "All generated seeds, raw lossless FLAC/MP3/codes and annotations retained in SQLite; never drop low-scoring or capped outputs",
  "updated_unix": 1794055964.185773,
  "content_enjoyment_head": "audiobox:CE",
  "production_quality_head": "audiobox:PQ",
  "genuineness_head": "genuineness_0_6"
}