{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-08-013",
    "source_story_id": "tmp-story-base-model-agent-prediction",
    "edition_id": "mp-2026-10-08-morning-0090",
    "edition_url": "https://themachinepress.com/edition/2026-10-08",
    "position": 13,
    "story_type": "dispatch",
    "section": "developer-tools",
    "editorial_classification": "editorial",
    "headline": "The Best Coding Base Model Could Be Ranked Before It Learned the Harness",
    "slug": "the-best-coding-base-model-could-be-ranked-before-it-learned-the-harness",
    "dek": "Three screens evaluated decisive actions inside successful trajectories instead of demanding cold-start tool use from an untrained checkpoint.",
    "summary": "Three screens evaluated decisive actions inside successful trajectories instead of demanding cold-start tool use from an untrained checkpoint.",
    "body_text": "Base checkpoints often fail an agent benchmark because they cannot yet format tool calls, hiding whether their weights contain useful coding behavior. This work replays successful post-training trajectories, finds the first cumulative patch that flips tests from failing to passing and evaluates base models at that decisive step. Bits-per-byte on the certified action, multiple-choice discrimination against verifier-rejected alternatives and prefix-conditioned pass-at-k all ranked ten base/post-trained model pairs in close agreement with eventual SWE-bench Verified performance. The screens still depend on successful trajectories and a verifier, but they may prevent expensive post-training runs on checkpoints whose apparent weakness is only harness fluency—or whose underlying coding distribution truly lacks the needed action.",
    "why_it_matters": "Three screens evaluated decisive actions inside successful trajectories instead of demanding cold-start tool use from an untrained checkpoint.",
    "limitations": [
      "Base checkpoints often fail an agent benchmark because they cannot yet format tool calls, hiding whether their weights contain useful coding behavior.",
      "The screens still depend on successful trajectories and a verifier, but they may prevent expensive post-training runs on checkpoints whose apparent weakness is only harness fluency—or whose underlying coding distribution truly lacks the needed action."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-08-013/the-best-coding-base-model-could-be-ranked-before-it-learned-the-harness",
    "json_url": "https://themachinepress.com/story/mp-2026-10-08-013.json",
    "first_published_at": "2026-10-08T09:00:00.000-04:00",
    "modified_at": "2026-10-08T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-08-013-001",
        "text": "Three screens evaluated decisive actions inside successful trajectories instead of demanding cold-start tool use from an untrained checkpoint.",
        "source_ids": [
          "source-2026-10-08-013"
        ],
        "qualification": "Base checkpoints often fail an agent benchmark because they cannot yet format tool calls, hiding whether their weights contain useful coding behavior."
      }
    ],
    "source_ids": [
      "source-2026-10-08-013"
    ],
    "tags": [
      "coding agents",
      "post-training",
      "SWE-bench"
    ],
    "image_url": "https://themachinepress.com/issues/2026-10-08/coding-agent-blue-screen.webp",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-08-013",
      "title": "arXiv preprint 2610.10478",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.10478",
      "canonical_url": "https://arxiv.org/abs/2610.10478",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-07T13:37:19.000-04:00",
      "accessed_at": "2026-10-08T08:26:48.909-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-08-013-001"
      ]
    }
  ],
  "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 Best Coding Base Model Could Be Ranked Before It Learned the Harness",
    "publisher": "The Machine Press",
    "published_at": "2026-10-08T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-08-013/the-best-coding-base-model-could-be-ranked-before-it-learned-the-harness"
  }
}