{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-05-007",
    "source_story_id": "tmp-story-failbench-robot-success-judging",
    "edition_id": "mp-2026-09-05-morning-0058",
    "edition_url": "https://themachinepress.com/edition/2026-09-05",
    "position": 7,
    "story_type": "dispatch",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "The Robot Judge Called Ambiguous Failures Success",
    "slug": "the-robot-judge-called-ambiguous-failures-success",
    "dek": "The best of thirteen vision-language detectors reached 0.77 balanced accuracy and fell near chance on contact-heavy assembly.",
    "summary": "The best of thirteen vision-language detectors reached 0.77 balanced accuracy and fell near chance on contact-heavy assembly.",
    "body_text": "FailBench brings together 2,197 manipulation attempts from fourteen public sources, with three-quarters of failures occurring naturally. General-purpose vision-language models beat systems fine-tuned specifically for failure detection, while the strongest detector reached 0.77 mean balanced accuracy. Performance approached saturation when object motion made outcomes visible but dropped below 0.60 on contact-intensive assembly tasks; ambiguous evidence systematically biased predictions toward success. Cropping outcome-relevant regions improved the top detector by 2.4 points without retraining.",
    "why_it_matters": "The best of thirteen vision-language detectors reached 0.77 balanced accuracy and fell near chance on contact-heavy assembly.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-05-007/the-robot-judge-called-ambiguous-failures-success",
    "json_url": "https://themachinepress.com/story/mp-2026-09-05-007.json",
    "first_published_at": "2026-09-05T09:00:00.000-04:00",
    "modified_at": "2026-09-05T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-05-007-001",
        "text": "The best of thirteen vision-language detectors reached 0.77 balanced accuracy and fell near chance on contact-heavy assembly.",
        "source_ids": [
          "source-2026-09-05-007"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-05-007"
    ],
    "tags": [
      "robot evaluation",
      "vision-language models",
      "failure detection"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-05-007",
      "title": "arXiv preprint 2609.03611",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.03611",
      "canonical_url": "https://arxiv.org/abs/2609.03611",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-03T05:58:55.000-04:00",
      "accessed_at": "2026-09-05T08:25:30.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-05-007-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Robot Judge Called Ambiguous Failures Success",
    "publisher": "The Machine Press",
    "published_at": "2026-09-05T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-05-007/the-robot-judge-called-ambiguous-failures-success"
  }
}