{
  "cases": 10,
  "repetitions": 3,
  "status": "Corrected-worker runs complete; approved for publication",
  "comparison": "Historical comparison; measurement dates and caller cohorts differ",
  "rerun_notes": [
    "Eleven results use three complete repetitions of each case after the buffered-speech watchdog fix. All 60 retained calls ended naturally.",
    "One entire Max pass interrupted by a scheduled deployment was repeated. Smaller sequential batches accommodated the remaining credit reservation; no cases were selected by score.",
    "One Max case was repeated after a protocol disconnect (code 1008). Its cause remains unresolved. The interrupted attempt scored 100%; the replay scored 75%. Retaining the interrupted attempt would put Max at 88.5%, rather than 87.7%.",
    "Dialt and Gemini retain the earlier two-minute-limit results. October 21 results use a three-minute limit. Unknown criteria contribute their midpoint."
  ],
  "arms": [
    {"label": "Dialt Fast", "score": 0.9245835, "p50_ms": 1088, "p90_ms": 1776, "group": "dialt", "date": "2026-10-05", "caller": "Irish Male"},
    {"label": "Dialt Smart", "score": 0.9944445, "p50_ms": 1501.5, "p90_ms": 2062, "group": "dialt", "date": "2026-10-05", "caller": "Irish Male"},
    {"label": "GPT Live-1", "score": 0.7967595, "p50_ms": 2273, "p90_ms": 3225, "group": "external", "date": "2026-10-21", "caller": "US Male", "limit_s": 180},
    {"label": "Gemini 3.8 Live", "score": 0.899, "p50_ms": 2018, "p90_ms": 2720, "group": "external", "date": "2026-09-26", "caller": "Irish Male"},
    {"label": "Eleven Realtime Mini", "score": 0.709074, "p50_ms": 2104.5, "p90_ms": 3544, "group": "eleven", "date": "2026-10-21", "caller": "US Male", "limit_s": 180},
    {"label": "Eleven Realtime Max", "score": 0.876852, "p50_ms": 2094, "p90_ms": 3395, "group": "eleven", "date": "2026-10-21", "caller": "US Male", "limit_s": 180}
  ]
}
