{
  "schema_version": "1.0",
  "id": "s13:https://arxiv.org/abs/2607.12962v1",
  "slug": "2607-12962v1-0q4i26c",
  "url": "https://feed7.dev/p/2607-12962v1-0q4i26c",
  "title": "Form, Not Content? A Preregistered, Placebo-Controlled Evaluation of Learned Error-Conditioned Self-Repair Through Prompts and Weights in Frozen Small Code Models",
  "why_included": "Placebo-controlled tests found no evidence that small frozen code models repaired failures because of the error content itself. Retry scaffolds and mismatched feedback performed as well or better.",
  "summary": "PoPE tested frozen **0.5–1.5B code models** with live errors and placebos through prompts and adapter training. In prompting, a form-only placebo unlocked **12 units versus 10** for live error patterns on a resistant 40-unit band.",
  "practical_implication": "When measuring self-repair, preserve the retry scaffold but remove or mismatch task-relevant feedback. If the placebo performs similarly, do not attribute gains to the compiler or test error; the retry format itself may be doing the work.",
  "agent_context": "PoPE tested frozen **0.5–1.5B code models** with live errors and placebos through prompts and adapter training. In prompting, a form-only placebo unlocked **12 units versus 10** for live error patterns on a resistant 40-unit band.\n\nWhen measuring self-repair, preserve the retry scaffold but remove or mismatch task-relevant feedback. If the placebo performs similarly, do not attribute gains to the compiler or test error; the retry format itself may be doing the work.\n\nIn adapter training, error content tied the no-intervention baseline at **8–8**, while a deranged-error placebo reached **10 unlocks**. These are public-tier screening results, not equivalence evidence; hidden-tier confirmation was deferred.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2607.12962v1",
    "published_at": "2026-07-14T16:59:42.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "benchmark",
  "domains": [
    "coding"
  ],
  "topics": [
    "agent-evals",
    "benchmark-integrity",
    "agent-reliability"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "In adapter training, error content tied the no-intervention baseline at **8–8**, while a deranged-error placebo reached **10 unlocks**. These are public-tier screening results, not equivalence evidence; hidden-tier confirmation was deferred."
  ],
  "lifecycle": "Current",
  "published_at": "2026-07-14T16:59:42.000Z",
  "modified_at": "2026-07-14T16:59:42.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2607-12962v1-0q4i26c",
    "json": "https://feed7.dev/p/2607-12962v1-0q4i26c.json",
    "markdown": "https://feed7.dev/p/2607-12962v1-0q4i26c.md"
  }
}