{
  "date": "2026-11-02",
  "repository": "https://github.com/GuHang2011/frozen-clip-fake-news-fusion",
  "commit": "f06dd9d3b48ecd2bb4e3c6882d36de1ceb8cb9b5",
  "command": "P2R=results python -I analysis/analyze3.py <output-directory>",
  "environment": {
    "python": "3.12.10",
    "numpy": "2.4.1",
    "platform": "Windows"
  },
  "scope": "Recompute few-shot per-run metrics from public saved hard predictions; no training or independent dataset reconstruction.",
  "datasets": [
    "gossipcop",
    "politifact"
  ],
  "shots": [
    "2",
    "4",
    "8",
    "16"
  ],
  "methods": [
    "FCP-fixed",
    "FCP-tuned",
    "LP(I)",
    "LP(T)",
    "LP(T+I)",
    "LP(T+I+K)",
    "MLP(T+I)",
    "MLP(T+I+K)",
    "Tip(T+I)"
  ],
  "seeds": [
    1,
    2,
    42
  ],
  "rows": 216,
  "metrics": [
    "pred_fake",
    "f1",
    "bacc"
  ],
  "absolute_tolerance_percentage_points": 1e-09,
  "max_absolute_difference_percentage_points": {
    "pred_fake": 0.0,
    "f1": 0.0,
    "bacc": 0.0
  },
  "reference_sha256": "da6c7575df0abb789325d461d41a77ecf881c247094ccc6dfc996305d05b56c8",
  "recomputed_sha256": "da6c7575df0abb789325d461d41a77ecf881c247094ccc6dfc996305d05b56c8",
  "passed": true,
  "not_run": [
    "Original training",
    "analyze.py bootstrap and aggregate analysis",
    "analyze2.py balanced-accuracy bootstrap",
    "Reconstruction of source data and original train/test splits"
  ],
  "limitations": [
    "Hard labels cannot reconstruct ranking-based AUROC, score calibration or threshold sweeps.",
    "Matching archived outputs checks this script and supplied artifacts, not independent sample alignment or causal validity.",
    "FCP-fixed and FCP-tuned are selection views; table rows are not new training runs."
  ]
}