{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-11-08-006",
    "source_story_id": "tmp-story-activation-inversion",
    "edition_id": "mp-2026-11-08-morning-0121",
    "edition_url": "https://themachinepress.com/edition/2026-11-08",
    "position": 6,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "A Model Writes the Text That Activates a Direction",
    "slug": "a-model-writes-the-text-that-activates-a-direction",
    "dek": "Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.",
    "summary": "Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.",
    "body_text": "Rather than searching a fixed corpus for text that activates a residual-stream direction, a metamodel generates candidate exemplars and optimizes their cosine similarity with the target activation. The reported Qwen3.6-27B system beat a ten-million-token corpus search on most held-out languages and domains, activated 89 percent of held-out sparse-autoencoder features, and recovered LoRA backdoor payloads more often than corpus search. Because a forward pass can verify each generated exemplar, the output remains testable, though an activating example is not automatically a complete semantic explanation.",
    "why_it_matters": "Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.",
    "limitations": [
      "Because a forward pass can verify each generated exemplar, the output remains testable, though an activating example is not automatically a complete semantic explanation."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-11-08-006/a-model-writes-the-text-that-activates-a-direction",
    "json_url": "https://themachinepress.com/story/mp-2026-11-08-006.json",
    "first_published_at": "2026-11-08T09:00:00.000-05:00",
    "modified_at": "2026-11-08T09:00:00.000-05:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-11-08-006-001",
        "text": "Maximally-Activating Example Metamodels turn residual-stream vectors into verifiable exemplars instead of waiting for a corpus hit.",
        "source_ids": [
          "source-2026-11-08-006"
        ],
        "qualification": "Because a forward pass can verify each generated exemplar, the output remains testable, though an activating example is not automatically a complete semantic explanation."
      }
    ],
    "source_ids": [
      "source-2026-11-08-006"
    ],
    "tags": [
      "mechanistic interpretability",
      "activation inversion",
      "model auditing"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-11-08-006",
      "title": "arXiv preprint 2611.08465, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2611.08465v1",
      "canonical_url": "https://arxiv.org/abs/2611.08465v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-11-05T13:54:40.000-05:00",
      "accessed_at": "2026-11-08T08:27:17.257-05:00",
      "supports_claim_ids": [
        "claim-mp-2026-11-08-006-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "A Model Writes the Text That Activates a Direction",
    "publisher": "The Machine Press",
    "published_at": "2026-11-08T09:00:00.000-05:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-11-08-006/a-model-writes-the-text-that-activates-a-direction"
  }
}