{
  "_comment": [
    "multitalker-parakeet-streaming-0.6b-v1 streaming per-tensor tolerances for validate_streaming.py.",
    "",
    "Calibration regime (matches /porting-2-oracle Stage 4 recipe):",
    "  - reference: NeMo 2.8.0rc0 conformer_stream_step on jfk.wav, R=13 (att_context [70,13]),",
    "    single-speaker EncDecMultiTalkerRNNTBPEModel (layer-0 spk-kernel injection applied).",
    "  - observed: C++ stream path, --backend cpu --threads 1, F32 GGUF, chunk cadence 500 ms.",
    "  - per kind = max-of-max across all chunks and dumped layers (0/12/23).",
    "  - per-entry: max_abs  = max(1.5 x observed_max_abs,  max(1e-4 x p99_abs, 1e-6))",
    "               mean_abs = max(1.5 x observed_mean_abs, max(1e-5 x rms,     1e-6))",
    "",
    "Model note: the streaming encoder applies the same always-on layer-0 speaker-kernel",
    "FF injection as the offline path (encoder.cpp::apply_spk_kernel_injection), per chunk",
    "on the post-drop block-0 input. Result: 150/150 tensors within tolerance, edit_dist 0,",
    "transcript byte-exact vs NeMo at R=13.",
    "",
    "Cross-check: streaming enc_out drift (obs 5.5e-6) is TIGHTER than the offline C++",
    "enc.final drift (1.9e-3) on the same audio -- clean per-chunk caches accumulate less",
    "than a single 24-block offline pass. Every kind's budget is magnitude-dominated except",
    "cache_lt / mel_in (1.5x-observed dominated). Same convention as the",
    "nemotron-speech-streaming-en-0.6b sibling.",
    "",
    "Pattern-keyed: stream.chunk.<N>.<kind>[_<layer>] all share the same '<kind>' entry;",
    "validate_streaming.py expands these at harness time."
  ],
  "cache_lc_in": {
    "max_abs": 1.8002947837114417e-05,
    "mean_abs": 1e-06,
    "_observed_max": 5.904585123062134e-06,
    "_observed_mean": 1.3644430385539863e-07,
    "_ref_p99": 0.18002947837114416,
    "_ref_rms": 0.06312107247075341
  },
  "cache_lc_out": {
    "max_abs": 1.801336298882962e-05,
    "mean_abs": 1e-06,
    "_observed_max": 5.904585123062134e-06,
    "_observed_mean": 1.3644430385539863e-07,
    "_ref_p99": 0.18013362988829618,
    "_ref_rms": 0.06470617190344634
  },
  "cache_lt_in": {
    "max_abs": 0.0020686249885559084,
    "mean_abs": 4.305163344313818e-05,
    "_observed_max": 0.00021599233150482178,
    "_observed_mean": 5.8020809245371685e-06,
    "_ref_p99": 20.686249885559082,
    "_ref_rms": 4.305163344313818
  },
  "cache_lt_out": {
    "max_abs": 0.0020686249885559084,
    "mean_abs": 4.305163344313818e-05,
    "_observed_max": 0.00021599233150482178,
    "_observed_mean": 5.8020809245371685e-06,
    "_ref_p99": 20.686249885559082,
    "_ref_rms": 4.305163344313818
  },
  "channel_len": {
    "max_abs": 0.007,
    "mean_abs": 0.0007000000000000001,
    "_observed_max": 0.0,
    "_observed_mean": 0.0,
    "_ref_p99": 70.0,
    "_ref_rms": 70.0
  },
  "enc_out": {
    "max_abs": 2.1179130151867858e-05,
    "mean_abs": 1e-06,
    "_observed_max": 5.4854899644851685e-06,
    "_observed_mean": 2.2248031176072152e-07,
    "_ref_p99": 0.21179130151867856,
    "_ref_rms": 0.07786172341954185
  },
  "mel_in": {
    "max_abs": 0.0016635532379150391,
    "mean_abs": 0.00011653641447048212,
    "_observed_max": 0.00025653839111328125,
    "_observed_mean": 2.6702478742453055e-06,
    "_ref_p99": 16.63553237915039,
    "_ref_rms": 11.65364144704821
  }
}