{
  "claim": "The auto-gpu-kernel report names four optimizations it converged on. Its committed experiment logs record 98 numbered experiments across the two DSA kernels, of which 32 were kept. Every row below is counted by hand from the two summary.md tables in archive/, classifying each experiment by the verdict its own Notes column gives.",
  "method": "Counted from archive/dsa_sparse_attention_h16_ckv512_kpe64_topk2048_ps64/experiments/summary.md (54 rows) and archive/dsa_topk_indexer_fp8_h64_d128_topk2048_ps64/experiments/summary.md (44 rows) at commit fetched 2026-09-18. \"Kept\" = the row is marked new best or an explicit marginal keep and is not later reverted by a subsequent row.",
  "source": "https://github.com/Dogacel/auto-gpu-kernel",
  "captured": "2026-09-18",
  "columns": [
    { "key": "track", "label": "kernel" },
    { "key": "logged", "label": "experiments logged", "align": "right" },
    { "key": "kept", "label": "kept", "align": "right" },
    { "key": "discarded", "label": "discarded", "align": "right" },
    { "key": "numerical", "label": "rejected by the correctness gate", "align": "right" },
    { "key": "compile", "label": "rejected by the compiler", "align": "right" },
    { "key": "rate", "label": "keep rate", "align": "right" }
  ],
  "rows": [
    {
      "track": "DSA sparse attention (23 traces)",
      "logged": 54,
      "kept": 17,
      "discarded": 37,
      "numerical": 1,
      "compile": 3,
      "rate": "31.5%"
    },
    {
      "track": "DSA top-k indexer (128 traces)",
      "logged": 44,
      "kept": 15,
      "discarded": 29,
      "numerical": 1,
      "compile": 0,
      "rate": "34.1%"
    },
    {
      "track": "both",
      "logged": 98,
      "kept": 32,
      "discarded": 66,
      "numerical": 2,
      "compile": 3,
      "rate": "32.7%"
    }
  ],
  "note": "The two correctness rejections are sparse-attention exp_27 (compact-block partition, quick run 1/2 passed, T=8 workload abs_err 2.82) and indexer exp_36 (skipping iteration 0 of the radix loop, 8/16 passed — negative learned weights make a negative final score, so the count above the initial threshold falls below topk). Both were reverted the same iteration. The three compiler rejections are sparse exp_20 (num_ctas=8 trips a Triton 3.6 CTA-planner assertion), exp_33 (BLOCK_N=256 needs 321 KB of shared memory) and exp_46 (.evict_last cannot be combined with .cg). Neither summary.md records an experiment that produced a latency number while failing the gate — the harness does not compute one."
}