{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-23-016",
    "source_story_id": "tmp-sidebar-tool-serving-stack-confounds",
    "edition_id": "mp-2026-09-23-morning-0076",
    "edition_url": "https://themachinepress.com/edition/2026-09-23",
    "position": 18,
    "story_type": "ticker",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "The Serving Stack Changed the Tool Score",
    "slug": "the-serving-stack-changed-the-tool-score",
    "dek": "Harness and serving behavior masqueraded as model failure in local evaluations.",
    "summary": "Harness and serving behavior masqueraded as model failure in local evaluations.",
    "body_text": "Identical tool-use requests behaved differently across Ollama, llama.cpp, vLLM and SGLang, and missing failure metadata could turn harness rejection into an apparent model non-call. Turn-pooled and per-instance estimates differed by as much as about 55 points.",
    "why_it_matters": "Harness and serving behavior masqueraded as model failure in local evaluations.",
    "limitations": [],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-23-016/the-serving-stack-changed-the-tool-score",
    "json_url": "https://themachinepress.com/story/mp-2026-09-23-016.json",
    "first_published_at": "2026-09-23T09:00:00.000-04:00",
    "modified_at": "2026-09-23T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-23-016-001",
        "text": "Harness and serving behavior masqueraded as model failure in local evaluations.",
        "source_ids": [
          "source-2026-09-23-018"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-23-018"
    ],
    "tags": [
      "tool use",
      "serving",
      "evaluation"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-23-018",
      "title": "arXiv preprint 2609.26693",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.26693",
      "canonical_url": "https://arxiv.org/abs/2609.26693",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-22T12:48:17.000-04:00",
      "accessed_at": "2026-09-23T08:22:43.661-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-23-016-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Serving Stack Changed the Tool Score",
    "publisher": "The Machine Press",
    "published_at": "2026-09-23T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-23-016/the-serving-stack-changed-the-tool-score"
  }
}