{
  "date": "2026-08-23",
  "device": "NVIDIA GB10",
  "cutlass_dsl": "4.6.2",
  "rows": 37810,
  "capacity": 2048,
  "chunk_count": 19,
  "final_chunk_rows": 946,
  "parity": {
    "blocks": [0, 24, 49],
    "roles": ["attn_qkv_proj", "attn_out_proj", "mlp_fc1"],
    "all_chunks_equal": true,
    "max_abs": 0.0,
    "checksums_equal": true
  },
  "timing_percent_improvement": {
    "block_0": {
      "attn_qkv_proj": 59.60311533167897,
      "attn_out_proj": 42.63660269340447,
      "mlp_fc1": 18.62094357187819
    },
    "block_24": {
      "attn_qkv_proj": 40.07175766590768,
      "attn_out_proj": -3.812088078043052,
      "mlp_fc1": 40.82958365166025
    },
    "block_49": {
      "attn_qkv_proj": 62.22094857579279,
      "attn_out_proj": 47.16679514736442,
      "mlp_fc1": -215.22616660569636
    }
  },
  "timing_interpretation": {
    "attn_qkv_proj": "Consistent 32 ms ring time and positive result across all blocks; advance to in-runtime validation.",
    "attn_out_proj": "Exact but timing is mixed; retain as experimental until in-runtime measurement.",
    "mlp_fc1": "Exact but duplicate-model unified-memory timing is unstable; no performance approval from this harness."
  },
  "benchmark_caveat": "The resident service and validator hold separate model copies on unified memory. Multi-gigabyte FC1 outputs caused pressure and one combined validator run stopped the container. Timing must be repeated inside the single resident model before deployment.",
  "policy": {
    "attn_qkv_proj": "advance",
    "attn_out_proj": "experimental",
    "mlp_fc1": "parity-approved, performance-pending",
    "mlp_fc2": "Comfy/cuBLAS fallback"
  },
  "next_gate": "Integrate only QKV behind an opt-in runtime flag, benchmark blocks 0, 24, and 49 in the resident model, then decide whether attention output or FC1 should be added."
}