{
  "schema": "anvil-serving.benchmark-workload/v1",
  "campaign_id": "2026-09-19-qwen38-huihui-ninfer",
  "selection": "Repository native deterministic preflight, intelligence/tool quality, and six-case synthetic image corpus. No evaluated-model self grading.",
  "launcher_revision": "2b99e3353eb2e5a2813195af85050b1da0754332",
  "files": [
    {
      "path": "anvil_serving/benchmarking/runner.py",
      "sha256": "f9fbf37f1a048a0bd7b02d36b1631754ee2df96f4654a4684d4f3e0e134549b8"
    },
    {
      "path": "anvil_serving/benchmarking/requests.py",
      "sha256": "20d43653e9f2938194be7b0ea111bc5206827bd0db37d18d071601e9490ea9a0"
    },
    {
      "path": "tests/fixtures/eval-data/agents-a1-v1/image-corpus.json",
      "sha256": "a514ae146b5aab3bcb55ae9fb27f55a41791a20863304d1bb95947e93931cd12"
    }
  ],
  "image_corpus": {
    "schema": "multimodal-corpus/v1",
    "license": "CC0-1.0",
    "media_hashes": "Each media SHA-256 is retained in the referenced repository corpus.",
    "cases": [
      "scene",
      "OCR",
      "chart",
      "UI",
      "spatial count",
      "multi-image comparison"
    ],
    "repetitions": 2,
    "total": 12
  },
  "quality": {
    "suites": [
      "intelligence",
      "tool"
    ],
    "checks": [
      "unified_diff_timeout_edit",
      "parallel_timeout_triage",
      "record_weather_zip typed zip string"
    ],
    "repetitions": 3,
    "required_pass_rate": 1.0,
    "candidate_off_visible_tokens": 1024,
    "candidate_low_visible_tokens": 1024,
    "candidate_low_reasoning_headroom": 4096,
    "incumbent_visible_tokens": 1024,
    "controls": "Direct loopback, sequential requests, C1. Thinking-off and reasoning-low are separate candidate cells; incumbent off is a correctness control."
  },
  "execution_order": [
    "initial smoke unsupported field failure",
    "smoke+JSON server default",
    "protocol preflight",
    "quality off",
    "quality reasoning-low",
    "vision-disabled failure",
    "managed vision reload and 12 image attempts",
    "managed unload then incumbent preflight and repeated quality",
    "managed unload and media restoration"
  ],
  "measurement": {
    "path": "online direct",
    "warm_cold_state": "Cold managed start followed by sequential mixed warm probes; no controlled performance series.",
    "statistics": "Pass counts only; incidental latency/throughput fields are not headline performance evidence."
  }
}