{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-22-017",
    "source_story_id": "tmp-brief-trainable-monitorability",
    "edition_id": "mp-2026-10-22-morning-0104",
    "edition_url": "https://themachinepress.com/edition/2026-10-22",
    "position": 19,
    "story_type": "ticker",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "Reasoning Monitorability Becomes a Training Target",
    "slug": "reasoning-monitorability-becomes-a-training-target",
    "dek": "An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.",
    "summary": "An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.",
    "body_text": "Post-training strengthened the connection between internal reasoning text and model outputs, improving judged monitorability and prediction of unsafe behavior in the reported tests. The result suggests monitorability can be optimized, but it does not make chain-of-thought a complete or tamper-proof oversight channel.",
    "why_it_matters": "An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.",
    "limitations": [
      "The result suggests monitorability can be optimized, but it does not make chain-of-thought a complete or tamper-proof oversight channel."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-22-017/reasoning-monitorability-becomes-a-training-target",
    "json_url": "https://themachinepress.com/story/mp-2026-10-22-017.json",
    "first_published_at": "2026-10-22T09:00:00.000-04:00",
    "modified_at": "2026-10-22T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-22-017-001",
        "text": "An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.",
        "source_ids": [
          "source-2026-10-22-019"
        ],
        "qualification": "The result suggests monitorability can be optimized, but it does not make chain-of-thought a complete or tamper-proof oversight channel."
      }
    ],
    "source_ids": [
      "source-2026-10-22-019"
    ],
    "tags": [
      "AI safety",
      "chain of thought",
      "monitoring"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-22-019",
      "title": "arXiv preprint 2610.29734, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.29734v1",
      "canonical_url": "https://arxiv.org/abs/2610.29734v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-21T09:35:57.000-04:00",
      "accessed_at": "2026-10-22T08:26:13.404-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-22-017-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Reasoning Monitorability Becomes a Training Target",
    "publisher": "The Machine Press",
    "published_at": "2026-10-22T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-22-017/reasoning-monitorability-becomes-a-training-target"
  }
}