{
  "baseline_revision": "4dba3b34f7af800c681f37ac83c22f0e83274f29",
  "candidate_revision": "a8b3ca4",
  "candidate_rounds": [
    "readpath-b1",
    "readpath-b2"
  ],
  "baseline_rounds": [
    "readpath-a1",
    "readpath-a2"
  ],
  "measured_responses_read_path": 72000,
  "rejected_experiment": {
    "rounds": [
      "final-a1",
      "final-b1",
      "final-b2",
      "final-a2"
    ],
    "measured_responses": 72000,
    "candidate_source": "candidate_revision plus rejected-combined-coverage.patch; never merged into implementation",
    "reason": "No broad improvement; filter latency regressed slightly after bounding facet memory."
  },
  "uncached_diagnostic": "uncached-old uses original baseline image; request_cache=false identically on all search endpoints. Counters before/after verify native request-cache bypass. Other engine caches remain enabled.",
  "payload_observation": "One representative benchmark point, compact payload JSON bytes; identical nested _source. Not response latency.",
  "ram_experiment": "One exploratory baseline-image round after on_disk_payload=false; optimizer green; configuration restored to true afterward. No production default change or speedup claim.",
  "summary_method": "median of all six trial medians across two rounds per variant, no fastest-trial selection",
  "host": "Shared Docker VM; no gateway builds during retained read-path or rejected-experiment timing. Not a dedicated performance host."
}