{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-17-006",
    "source_story_id": "tmp-story-frontier-model-safety-evidence",
    "edition_id": "mp-2026-09-17-morning-0070",
    "edition_url": "https://themachinepress.com/edition/2026-09-17",
    "position": 6,
    "story_type": "dispatch",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "Safety Evidence Lost to Retrieval Friction",
    "slug": "safety-evidence-lost-to-retrieval-friction",
    "dek": "Four frontier models reacted strongly to severity and cost, while a stated risk jump from 10% to 70% moved inspection by at most 21 points.",
    "summary": "Four frontier models reacted strongly to severity and cost, while a stated risk jump from 10% to 70% moved inspection by at most 21 points.",
    "body_text": "SAFE tests whether a model asks for safety-relevant evidence before making a deployment decision. Across GPT-5.5, o3, Claude Opus 4.8 and Claude Sonnet 4.6, inspection increased with severity and fell with retrieval cost. Stated probability mattered less: moving a problem from 10% to 70% likelihood changed inspection by no more than 21 percentage points. Opus inspected almost by default; o3 skipped most and reacted most sharply to thresholds. Counterfactual framing changed some decisions without appearing in the explanations, exposing a gap between rationales and acquisition policy.",
    "why_it_matters": "Four frontier models reacted strongly to severity and cost, while a stated risk jump from 10% to 70% moved inspection by at most 21 points.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-17-006/safety-evidence-lost-to-retrieval-friction",
    "json_url": "https://themachinepress.com/story/mp-2026-09-17-006.json",
    "first_published_at": "2026-09-17T09:00:00.000-04:00",
    "modified_at": "2026-09-17T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-17-006-001",
        "text": "Four frontier models reacted strongly to severity and cost, while a stated risk jump from 10% to 70% moved inspection by at most 21 points.",
        "source_ids": [
          "source-2026-09-17-006"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-17-006"
    ],
    "tags": [
      "AI safety",
      "evidence acquisition",
      "model behavior"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-17-006",
      "title": "arXiv preprint 2609.17865",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.17865",
      "canonical_url": "https://arxiv.org/abs/2609.17865",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-15T17:45:43.000-04:00",
      "accessed_at": "2026-09-17T08:28:34.205-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-17-006-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Safety Evidence Lost to Retrieval Friction",
    "publisher": "The Machine Press",
    "published_at": "2026-09-17T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-17-006/safety-evidence-lost-to-retrieval-friction"
  }
}