{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-09-016",
    "source_story_id": "tmp-sidebar-circuit-anchored-evolution",
    "edition_id": "mp-2026-08-09-morning-0031",
    "edition_url": "https://themachinepress.com/edition/2026-08-09",
    "position": 18,
    "story_type": "ticker",
    "section": "safety",
    "editorial_classification": "editorial",
    "headline": "Two Percent of Features Held the Safety Line",
    "slug": "two-percent-of-features-held-the-safety-line",
    "dek": "Anchoring a small identified circuit preserved safety during model self-evolution with limited capability cost.",
    "summary": "Anchoring a small identified circuit preserved safety during model self-evolution with limited capability cost.",
    "body_text": "Across three model families and two evolution algorithms, the authors report that constraining an identified safety circuit representing under two percent of features preserved safety better than explicit reward constraints. The result is model- and evaluation-dependent evidence for circuit anchoring, not a general guarantee against unsafe adaptation.",
    "why_it_matters": "Anchoring a small identified circuit preserved safety during model self-evolution with limited capability cost.",
    "limitations": [
      "The result is model- and evaluation-dependent evidence for circuit anchoring, not a general guarantee against unsafe adaptation."
    ],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-09-016/two-percent-of-features-held-the-safety-line",
    "json_url": "https://themachinepress.com/story/mp-2026-08-09-016.json",
    "first_published_at": "2026-08-09T09:00:00.000-04:00",
    "modified_at": "2026-08-09T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-09-016-001",
        "text": "Anchoring a small identified circuit preserved safety during model self-evolution with limited capability cost.",
        "source_ids": [
          "source-2026-08-09-018"
        ],
        "qualification": "The result is model- and evaluation-dependent evidence for circuit anchoring, not a general guarantee against unsafe adaptation."
      }
    ],
    "source_ids": [
      "source-2026-08-09-018"
    ],
    "tags": [
      "model evolution",
      "mechanistic interpretability",
      "safety"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-09-018",
      "title": "arXiv preprint 2608.05158",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.05158",
      "canonical_url": "https://arxiv.org/abs/2608.05158",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": null,
      "accessed_at": "2026-08-09T08:17:45.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-09-016-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Two Percent of Features Held the Safety Line",
    "publisher": "The Machine Press",
    "published_at": "2026-08-09T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-09-016/two-percent-of-features-held-the-safety-line"
  }
}