{
  "_comment": [
    "Finalized Stage-4 tolerances for MOSS-Transcribe-Diarize (moss).",
    "CORRECTNESS REGIME: reference-dtype BF16 GGUF; KV cache F16; reference mel",
    "(C++ MelFrontend, validated separately in Step 7); backend cpu, threads 1",
    "(validate.py defaults). Every number below is measured in this regime via",
    "`validate.py all` and finalized as max(1.5 x observed C++ drift, Stage-2",
    "provisional magnitude budget, 1e-6).",
    "DOMINANT DRIFT SOURCE: the reference runs BF16 (torch, attn_implementation",
    "eager); the C++ path dequantizes BF16 weights to F32 and computes in F32",
    "(F16 conv kernels with F32 accumulation). So C++ is MORE precise than the",
    "reference and the gap is bf16-reference storage/compute vs f32 C++ - a",
    "constant ~1-3% relative drift, not a bug. Confirmed WER-neutral: on the",
    "first 100 LibriSpeech test-clean utterances the C++ ref-dtype WER (1.40%)",
    "is bit-identical to the Oracle reference WER on the same subset (1.40%).",
    "LARGE ABSOLUTE MAXES ARE EXPECTED, NOT BUGS: the Whisper encoder and Qwen3",
    "decoder residual streams carry genuinely large pre-normalization activations",
    "(reference |enc.block.23| reaches ~3.6e3, |dec.out_before_head| ~1e2). bf16",
    "relative error on those single largest elements yields the big max_abs here;",
    "the following LayerNorm/RMSNorm renormalizes them (enc.ln_post drift falls",
    "back to ~1.8e-3 mean). The per-tensor MEAN budgets are the meaningful gate.",
    "ENCODER PADDING: the encoder runs on the reference's 30s zero-padded chunk;",
    "frames beyond token_len*4 (jfk: 552 of 1500) are ill-conditioned and TRIMMED",
    "before the 4x merge + adaptor, so they never reach the decoder. Localized",
    "host-side: real-region (0:552) and padding-region drift are the same order;",
    "the huge maxes sit on large-magnitude activations, not a mask/shape bug.",
    "dec.token_emb is a pure get_rows of the GGUF embedding table -> pinned exact",
    "0.0 (zero-drift exception). No _provisional flags remain."
  ],
  "dec.token_emb": {
    "max_abs": 0.0,
    "mean_abs": 0.0
  },
  "enc.mel.in": {
    "max_abs": 9.871853202581411e-05,
    "mean_abs": 5.1495853070020074e-06,
    "_comment": "C++ MelFrontend vs reference WhisperFeatureExtractor; observed drift (2.2e-5/3.4e-8) is below the Stage-2 magnitude budget, kept."
  },
  "enc.pos_add.out": {
    "max_abs": 0.05,
    "mean_abs": 0.0022
  },
  "enc.block.0.out": {
    "max_abs": 0.12,
    "mean_abs": 0.004
  },
  "enc.block.23.out": {
    "max_abs": 3100.0,
    "mean_abs": 0.08,
    "_comment": "pre-final-LN residual; ref |max| ~3.6e3, bf16 rel error dominates max_abs. Normalized away by enc.ln_post; mean is the real gate."
  },
  "enc.ln_post.out": {
    "max_abs": 13.0,
    "mean_abs": 0.003
  },
  "enc.merge.out": {
    "max_abs": 1.5,
    "mean_abs": 0.003
  },
  "enc.adaptor.out": {
    "max_abs": 0.4,
    "mean_abs": 0.012,
    "_comment": "decoder audio handoff (post-LayerNorm); rel_mean 0.96% - clean bf16 precision."
  },
  "dec.audio_injected": {
    "max_abs": 0.4,
    "mean_abs": 0.007
  },
  "dec.block.0.out": {
    "max_abs": 0.75,
    "mean_abs": 0.0095
  },
  "dec.block.27.out": {
    "max_abs": 320.0,
    "mean_abs": 0.26,
    "_comment": "pre-final-RMSNorm residual; large activations, bf16 accumulation over 28 layers. Renormalized before the head."
  },
  "dec.out_before_head": {
    "max_abs": 8.5,
    "mean_abs": 0.08
  },
  "dec.logits_raw": {
    "max_abs": 0.48,
    "mean_abs": 0.072,
    "_comment": "first-step (prefill) logits; argmax preserved (transcript exact) despite bf16 drift."
  },
  "dec.logits_raw.gen8": {
    "max_abs": 0.5,
    "mean_abs": 0.065,
    "_comment": "mid-generation (n_past>0) logits at greedy step 8 - KV-cache decode coverage."
  }
}