{
  "what": "Re-judge of the 140-round quickstart log with the decision panel at --concurrency 8, at evaluatorq's default cap of 10 in-flight judge calls and with the cap raised to 48. Wall time is the whole process (uv startup included). A lost vote is a judge vote with completed=false; every one here was pplx-decider-v1-27b answering 429. Decided rounds are the CLI's 'decisive' count.",
  "date": "2026-10-21",
  "evaluatorq": "1.74.0",
  "arena_commit": "3f5a894",
  "command_cap10": "orq-arena rejudge examples/quickstart/battles.jsonl --config docs/research/configs/quickstart_decisions.yaml --judge-api decisions --judge typesafe/jev-latest --judge perplexity/pplx-decider-v1-27b --judge openai/gpt-6-luna --concurrency 8 --output OUT.jsonl --report-json OUT.json",
  "command_cap48": "uv run python scripts/rejudge_cap.py 48 rejudge ...same arguments as command_cap10 (runs 3 and 4 used the script for both caps)",
  "count_lost_votes": "sum(not v['completed'] for line in open('OUT.jsonl') for v in json.loads(line)['judge_votes'])",
  "runs": [
    {
      "order": 1,
      "cap": 10,
      "wall_s": 35,
      "lost_votes": 6,
      "decided_rounds": 109
    },
    {
      "order": 2,
      "cap": 48,
      "wall_s": 41,
      "lost_votes": 57,
      "decided_rounds": 93
    },
    {
      "order": 3,
      "cap": 48,
      "wall_s": 44,
      "lost_votes": 62,
      "decided_rounds": 95
    },
    {
      "order": 4,
      "cap": 10,
      "wall_s": 48,
      "lost_votes": 4,
      "decided_rounds": 113
    }
  ],
  "notes": "Runs 1 and 2 ran back to back; runs 3 and 4 in reverse order with 90 s pauses before each, so the rate-limit window from a previous run does not explain the gap."
}
