{
  "schema_version": 1,
  "adapter_key": "video-opsd-checkpoint",
  "validated_at": "2026-09-01",
  "accelerator": {"vendor": "NVIDIA", "model": "A100"},
  "command": [
    "auto-research", "video-opsd-eval",
    "--annotations", "docs/gpu-validations/fixtures/video-opsd-demo-evidence.jsonl",
    "--video-root", "data/video-mme-v2-demo",
    "--model-id", "HuggingFaceTB/SmolVLM2-256M-Video-Instruct",
    "--model-revision", "067788b187b95ebe7b2e040b3e4299e342e5b8fd",
    "--maximum-examples", "1", "--num-frames", "32",
    "--maximum-new-tokens", "12", "--seeds", "42,43,44"
  ],
  "dataset": {
    "name": "MME-Benchmarks/Video-MME-v2 official demo asset",
    "revision": "6e4bebb03202e1ddbf3d37703e560e51c5aa2d64",
    "asset": "assets/demo.mp4",
    "questions": 1,
    "annotation": "manual protocol evidence frames; not an upstream paper annotation"
  },
  "checkpoint": {
    "model_id": "HuggingFaceTB/SmolVLM2-256M-Video-Instruct",
    "revision": "067788b187b95ebe7b2e040b3e4299e342e5b8fd"
  },
  "result": "passed",
  "metrics": {
    "seeds": [42, 43, 44],
    "full_video_accuracy_mean": 0.0,
    "evidence_view_accuracy_mean": 1.0,
    "answer_agreement_mean": 0.0,
    "full_video_parse_rate_mean": 1.0,
    "evidence_view_parse_rate_mean": 1.0,
    "claim_boundary": "single official demo protocol smoke; not the paper training result"
  },
  "provenance": {
    "commit": "583e7a439003e6d4581d2415bd678289553b0e7e",
    "artifact_path": "docs/gpu-validations/video-opsd-checkpoint-a100-20260901.json",
    "raw_predictions_committed": false,
    "checkpoint_committed": false
  }
}