{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-15-017",
    "source_story_id": "tmp-brief-manaca-ptbr-safety-alignment",
    "edition_id": "mp-2026-10-15-morning-0097",
    "edition_url": "https://themachinepress.com/edition/2026-10-15",
    "position": 19,
    "story_type": "ticker",
    "section": "safety",
    "editorial_classification": "editorial",
    "headline": "A Safety Reward Reaches One While the Model Does Not Move",
    "slug": "a-safety-reward-reaches-one-while-the-model-does-not-move",
    "dek": "The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.",
    "summary": "The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.",
    "body_text": "For its 1.72-billion-parameter Brazilian-Portuguese model, the team says dedicated safety SFT—not DPO—moved the safety-helpfulness frontier. It reports a 94.1 percent lower safety bound above four compared baselines and releases the supporting code, data and logs.",
    "why_it_matters": "The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.",
    "limitations": [
      "For its 1.72-billion-parameter Brazilian-Portuguese model, the team says dedicated safety SFT—not DPO—moved the safety-helpfulness frontier."
    ],
    "importance": 6,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-15-017/a-safety-reward-reaches-one-while-the-model-does-not-move",
    "json_url": "https://themachinepress.com/story/mp-2026-10-15-017.json",
    "first_published_at": "2026-10-15T09:00:00.000-04:00",
    "modified_at": "2026-10-15T09:00:00.000-04:00",
    "content_status": "updated",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-15-017-001",
        "text": "The Manacá team documents three failed DPO attempts, a non-monotonic over-refusal curve and an LLM judge that silently refuses some evaluations.",
        "source_ids": [
          "source-2026-10-15-019"
        ],
        "qualification": "For its 1.72-billion-parameter Brazilian-Portuguese model, the team says dedicated safety SFT—not DPO—moved the safety-helpfulness frontier."
      }
    ],
    "source_ids": [
      "source-2026-10-15-019"
    ],
    "tags": [
      "Brazilian Portuguese",
      "alignment",
      "evaluation"
    ],
    "image_url": null,
    "corrections": [
      {
        "correction_id": "correction-2026-10-15-001",
        "revision": 2,
        "summary": "Publication note: This October 15 edition was prepared before its scheduled 9:00 AM America/Detroit reveal, but publication was delayed by a production build failure. The original reporting and image-generation timestamps are preserved.",
        "published_at": "2026-10-15T09:00:00.000-04:00",
        "affected_story_ids": [],
        "affected_claim_ids": []
      }
    ]
  },
  "sources": [
    {
      "source_id": "source-2026-10-15-019",
      "title": "arXiv preprint 2610.20570, original v1",
      "publisher": "arXiv / submitting research authors",
      "url": "https://arxiv.org/abs/2610.20570v1",
      "canonical_url": "https://arxiv.org/abs/2610.20570v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-14T13:53:46.000-04:00",
      "accessed_at": "2026-10-15T08:28:29.903-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-15-017-001"
      ]
    }
  ],
  "corrections": [
    {
      "correction_id": "correction-2026-10-15-001",
      "revision": 2,
      "summary": "Publication note: This October 15 edition was prepared before its scheduled 9:00 AM America/Detroit reveal, but publication was delayed by a production build failure. The original reporting and image-generation timestamps are preserved.",
      "published_at": "2026-10-15T09:00:00.000-04:00",
      "affected_story_ids": [],
      "affected_claim_ids": []
    }
  ],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "A Safety Reward Reaches One While the Model Does Not Move",
    "publisher": "The Machine Press",
    "published_at": "2026-10-15T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-15-017/a-safety-reward-reaches-one-while-the-model-does-not-move"
  }
}