{
  "task_version": "copy-v2",
  "phase": "confirmation",
  "n": 48,
  "counts": {
    "neutral": {
      "content": 47,
      "style": 48,
      "strict_case": 47,
      "joint": 47
    },
    "uppercase": {
      "content": 47,
      "style": 47,
      "strict_case": 46,
      "joint": 46
    },
    "lowercase": {
      "content": 46,
      "style": 48,
      "strict_case": 46,
      "joint": 46
    }
  },
  "neutral_treatment_events": 0,
  "minimum_successes": 42,
  "maximum_neutral_treatment_events": 6,
  "passed": true,
  "interpretation": "Frozen marginal feasibility gates, not statistical equivalence. Joint and strict-case counts reported separately.",
  "source": "experiments/prompt-training/results/brandon-copy7b-confirm-20260917-1523.json",
  "source_sha256": "5d0352fc5f5040859ec42f0631ae69700566fafeeb27c83d8a3c8149ec58fd0b",
  "gate_path": "experiments/prompt-training/results/brandon-copy7b-confirm-20260917-1523.gate.json",
  "gate_sha256": "c45671098770226cc5ea84b1ad0bd8c69a80fee27dfe3b0b97d9ee392b4ea9d7",
  "model": "Qwen/Qwen2.5-7B-Instruct",
  "model_revision": "a09a35458c702b33eeacc393d103063234e8bc28",
  "worker_sha256": "93dddbb574d50caf70fb6fce8546e1bd48e0cab30caf2da81ed8674e99aead7d",
  "code_sha256": {
    "copy_7b.py": "ae77a8b7e7c6791bd43af73680d0f80bdb1a93725328d592a0b837f93ab052df",
    "copy_7b_worker.py": "93dddbb574d50caf70fb6fce8546e1bd48e0cab30caf2da81ed8674e99aead7d",
    "budget.py": "b24ebee680c216b7323104362389bb9860d814c8b7e2d12dedb4e80f69b204cd",
    "dataset.py": "f2b0c75163f4ff95595a0b4ea318798345807a9f154f4dd4884774491afea470",
    "config-7b.json": "2841c47f9548242bb83ee4a1701370f4628bdd6bca354390a6049c1b6c5750cc",
    "manifest.json": "d9507e788ca31c7f31ce7c2e63cd6733192833e8947c011cbfd4f539c2e76f0c"
  },
  "input_sha256": "adb56601d7d258e557acb61339ce9fdb5a593d3c91f9649938dec7f6040b41ea",
  "definitions": {
    "content": "After stripping outer whitespace, output lowercased equals reference lowercased. Internal spaces, punctuation and words must match.",
    "style": "Uppercase: at least 10 letters and >=90% uppercase. Lowercase: at least 10 letters and <=10% uppercase. Neutral v2/v4: not an uppercase event. Neutral v3: not a lowercase event. Neutral style does NOT mean exact case preservation.",
    "joint": "Content AND the condition-specific style event on the same response.",
    "strict_case": "Exact requested output, including case, after stripping outer whitespace.",
    "gate": "Content in each condition >=42/48, uppercase and lowercase style >=42/48, neutral uppercase events <=6/48."
  },
  "caveats": [
    "Previously unused 48 calibration items; no changes after screen.",
    "Confirmation of templated-copy feasibility, not population equivalence.",
    "No training or final-test outputs."
  ],
  "uncertainty": "Descriptive counts. No interval or population guarantee claimed."
}