{
  "study": "Sleep Agent preliminary overnight comparison",
  "date": "2026-11-08",
  "model": "Qwen3.5-4B",
  "seed": 42,
  "decoding": "greedy",
  "patch_output_budget": 4096,
  "control_output_budget": 384,
  "primary_tasks": 3,
  "within_repository_transfer_tasks": 4,
  "control_questions": 40,
  "fresh_responses": 235,
  "methods": [
    {
      "method": "baseline",
      "primary_solved": 0,
      "transfer_solved": 0,
      "accepted_edits": 6,
      "control_correct": 17,
      "gained_vs_baseline": 0,
      "lost_vs_baseline": 0
    },
    {
      "method": "reference_sft",
      "primary_solved": 0,
      "transfer_solved": 0,
      "accepted_edits": 2,
      "control_correct": 21,
      "gained_vs_baseline": 4,
      "lost_vs_baseline": 0
    },
    {
      "method": "own_experience",
      "primary_solved": 0,
      "transfer_solved": 0,
      "accepted_edits": 1,
      "control_correct": 24,
      "gained_vs_baseline": 9,
      "lost_vs_baseline": 2
    },
    {
      "method": "other_experience",
      "primary_solved": 0,
      "transfer_solved": 0,
      "accepted_edits": 0,
      "control_correct": 24,
      "gained_vs_baseline": 8,
      "lost_vs_baseline": 1
    },
    {
      "method": "rag",
      "primary_solved": 0,
      "transfer_solved": 0,
      "accepted_edits": 0,
      "control_correct": 18,
      "gained_vs_baseline": 5,
      "lost_vs_baseline": 4
    }
  ],
  "limitations": [
    "One seed, one repository, seven coding tasks.",
    "Control questions came from prior development material.",
    "The reduced output budget was chosen after observing earlier runs.",
    "This summary is not a self-contained reproduction package."
  ]
}
