{
  "artifact": "s06_weka_prefix_hit_rate",
  "sprint": "S06-C3",
  "claim": "SemiAnalysis's published 96.57% local prefix-cache hit rate on cc-traces-weka-042026, recomputed from the dataset's own 64-token hash_ids under its stated definition and under the weaker any-seen definition.",
  "method": "per trace, per request: (a) contiguous common prefix with the previous request; (b) blocks previously seen in the trace; turn 0 counts 0 hits; aggregate over all blocks of all traces.",
  "prereg": "results/s06_weka_prefix_hit_rate.prereg.json",
  "source": {
    "dataset": "semianalysisai/cc-traces-weka-042026",
    "file": "traces.jsonl",
    "sha256": "d9e30ba2120d624cb5818a7b19f0150b1181b18ba2e4fe0c05c64218713eed33",
    "license": "Apache-2.0",
    "block_tokens": 64,
    "hash_id_scope": "local"
  },
  "published": {
    "hit_rate": 0.9657,
    "total_blocks": 129409824
  },
  "recomputed": {
    "total_blocks": 129409824,
    "hit_rate_a_prefix_vs_previous": 0.9657367202663069,
    "hit_rate_b_any_seen": 0.9678318316853596,
    "n_requests": 59204,
    "per_request_a_median": 0.9934949733885275,
    "n_requests_without_hash_ids": 70
  },
  "total_blocks_match_published": true,
  "abs_diff_a": 3.7e-05,
  "abs_diff_b": 0.002132,
  "tolerance": 0.005,
  "controls": {
    "fresh_ids": {
      "total_blocks": 129409824,
      "hit_rate_a_prefix_vs_previous": 0.0,
      "hit_rate_b_any_seen": 0.0,
      "n_requests": 59204,
      "per_request_a_median": 0.0,
      "n_requests_without_hash_ids": 70,
      "ok": true
    },
    "shuffled_within_request": {
      "total_blocks": 129409824,
      "hit_rate_a_prefix_vs_previous": 2.8591337857008443e-07,
      "hit_rate_b_any_seen": 0.9678318316853596,
      "n_requests": 59204,
      "per_request_a_median": 0.0,
      "n_requests_without_hash_ids": 70,
      "ok": true
    }
  },
  "verdict": "REPRODUCES under the stated definition (a): published 0.9657 over 129,409,824 blocks; recomputed (a) prefix-vs-previous 0.9657, (b) any-seen 0.9678, over 129,409,824 blocks in 739 traces / 59,204 requests. Controls: fresh ids -> 0/0 as required; shuffled -> (a) 0.0000, (b) unchanged OK.",
  "wall_s": 39.8,
  "cost_usd": 0.0,
  "does_not_prove": [
    "A per-trace hash scope: cross-session sharing (the fleet question) is not measurable on this corpus.",
    "A hit rate over blocks, not tokens or dollars; turn-0 blocks are counted as misses exactly as the publisher did.",
    "Nothing about the engine: this is arithmetic on the publisher's ids, which is why it can grade the publisher's number."
  ],
  "generated_utc": "2026-09-02T22:52:55.128003+00:00",
  "git_rev": "575776c",
  "git_dirty": true,
  "run_id": "59d465c3d4d44e79",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
}