{
  "expected": {
    "accept_all_p95_under_threshold_at_every_load": false,
    "constitution_p95_under_threshold_at_every_load": true,
    "n_arms": 5,
    "n_control_rows": 1750,
    "n_legitimate_rows": 19650,
    "n_loads": 14,
    "n_overload_rows": 78600,
    "n_seeds": 5,
    "threshold_legitimate_p95_us": 25000.0
  },
  "hash_version": 2,
  "name": "ap_resource_tail_latency",
  "note": "S06 row 3, sealed 2026-09-02 BEFORE tools/check_ap_resource_tail_latency.py existed. Question the dataset artifacts/native/ap_resource_gate_events.jsonl (100,000 decision rows + 350 finish + 1 header) was never asked: the legitimate-client p50/p95/p99 end-to-end latency (response.e2e_us) per arm x legitimate_load x seed, judged against the header's own thresholds.legitimate_p95_us = 25000. RECORDED (from the header and row counts): n_arms, n_loads, n_seeds, the three score_class counts, the threshold. PREDICTED: constitution under the bar at every load; accept_all over it at some load. BLIND (not in expected, reported whatever they are): token_bucket, station_quota, class_fair_share verdicts; the p99 values; the per-load first crossing. Negative control: a synthetic dataset whose legitimate rows carry e2e_us = 30000 must make constitution FALSE. Kill: constitution misses 25,000 us at any load -> publish the loads where its own bar fails.",
  "note_sha256": "a0d6f92e5d5c5742f0eb29c0578d1e130c5a44ee4bf7a12ad1460edd3a76d8cb",
  "protocol": "pre-registered before the benchmark run. sha256 covers expected+tol; note_sha256 covers the prose. Neither proves ordering on its own -- the git commit of this file does, and git_blob_matches() is the anchor that makes an edit visible.",
  "sealed": true,
  "sha256": "44d1b0a01594d14345c52bbff469f836e1bd2db051f7601e88ecad70aa0d9f1e",
  "tol": {}
}
