{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-07-007",
    "source_story_id": "tmp-story-perfreasoning-hardware-modeling",
    "edition_id": "mp-2026-09-07-morning-0060",
    "edition_url": "https://themachinepress.com/edition/2026-09-07",
    "position": 7,
    "story_type": "dispatch",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "Models Could Answer the Hardware Question. Most Could Not Build the Model",
    "slug": "models-could-answer-the-hardware-question-most-could-not-build-the-model",
    "dek": "Top systems cleared 90 percent on performance reasoning, yet nearly every configuration averaged below 15 percent when asked to generate analytical model code.",
    "summary": "Top systems cleared 90 percent on performance reasoning, yet nearly every configuration averaged below 15 percent when asked to generate analytical model code.",
    "body_text": "PerfReasoning tests whether language models can compare workload mappings, predict off-chip traffic and calculate buffer requirements. The strongest closed models exceeded 90 percent on reasoning questions and the best open-weight model reached 82.4 percent. Model construction was much harder: one reported configuration exceeded 80 percent, while all others averaged below 15 percent and varied across runs. The benchmark exposes a gap between plausible architectural answers and executable performance models; it does not measure every form of systems engineering.",
    "why_it_matters": "Top systems cleared 90 percent on performance reasoning, yet nearly every configuration averaged below 15 percent when asked to generate analytical model code.",
    "limitations": [
      "The benchmark exposes a gap between plausible architectural answers and executable performance models; it does not measure every form of systems engineering."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-07-007/models-could-answer-the-hardware-question-most-could-not-build-the-model",
    "json_url": "https://themachinepress.com/story/mp-2026-09-07-007.json",
    "first_published_at": "2026-09-07T09:00:00.000-04:00",
    "modified_at": "2026-09-07T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-07-007-001",
        "text": "Top systems cleared 90 percent on performance reasoning, yet nearly every configuration averaged below 15 percent when asked to generate analytical model code.",
        "source_ids": [
          "source-2026-09-07-007"
        ],
        "qualification": "The benchmark exposes a gap between plausible architectural answers and executable performance models; it does not measure every form of systems engineering."
      }
    ],
    "source_ids": [
      "source-2026-09-07-007"
    ],
    "tags": [
      "hardware performance",
      "code generation",
      "benchmark"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-07-007",
      "title": "arXiv preprint 2609.04476",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.04476",
      "canonical_url": "https://arxiv.org/abs/2609.04476",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-03T17:00:06.000-04:00",
      "accessed_at": "2026-09-07T08:20:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-07-007-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Models Could Answer the Hardware Question. Most Could Not Build the Model",
    "publisher": "The Machine Press",
    "published_at": "2026-09-07T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-07-007/models-could-answer-the-hardware-question-most-could-not-build-the-model"
  }
}