{
  "distinct_mpp_tiles": 72,
  "attention_layout_settings": 6,
  "unpack_settings": 12,
  "decisions": [
    {
      "id": "attention-layout",
      "decision": "rejected",
      "evidence": "screen.json",
      "reason": "All six preserve bits. Best layout gain below 0.5%; full QKV copying slower; no dispatch change."
    },
    {
      "id": "vector-unpack",
      "decision": "rejected-for-throughput",
      "evidence": "screen.json",
      "reason": "Decode alone improves about 20%, but decode+GEMM benefits are small or negative. Scalar codec remains the deployed implementation."
    },
    {
      "id": "dual-fc1-swiglu",
      "decision": "rejected",
      "evidence": "fc1.json",
      "reason": "Eight tested tiles preserve bits, all slower than compiled native FC1+SwiGLU. Best reduction_pct=-0.6940844599806928"
    },
    {
      "id": "bk2048",
      "decision": "rejected",
      "evidence": "panel.json",
      "reason": "Operator improvement did not survive the complete exact fused block; reduction_pct=-0.1895493265225756"
    },
    {
      "id": "fc2-range",
      "decision": "qualified-operator",
      "evidence": [
        "shapes.json",
        "shapes-extra.json"
      ],
      "reason": "Positive exact gains for real product and edit activations and repeated real edit activations at 8192. Under 1% at 16478, so native 4K fallback retained."
    },
    {
      "id": "norm-tg256-512",
      "decision": "rejected-quality",
      "evidence": "norm.json",
      "reason": "Changed BF16 values even with native rsqrt and four-lane reads. No deployment or full-image speed claim; previous differing-normalization image stacks remain in v2."
    },
    {
      "id": "norm-tg1024",
      "decision": "qualified-operator-and-block",
      "evidence": "norm.json",
      "reason": "Native loop/reduction order and BF16 boundaries match on actual first-block operator and complete block; full production trajectory validation required."
    },
    {
      "id": "qk-native-reduction",
      "decision": "qualified-full-research-suite",
      "evidence": [
        "blocks.json",
        "research/quality-audit.json"
      ],
      "reason": "Four contiguous reads per lane + precise rsqrt correct the earlier differing QK fusion; exact full tensors and pixels on four cases in research runner."
    },
    {
      "id": "selected-embedding-rows",
      "decision": "accepted",
      "evidence": "embedding/summary.json",
      "reason": "Native selected BF16 payload bits match across fresh-process AB/BA trials, padding/duplicates/endpoints covered; phase-specific improvement only."
    }
  ]
}