A checkable research artifact

01 / What was verified

216 per-run records matched the repository reference across three reported metrics. The check used public saved predictions; it did not train models.

Rows checked

216

Datasets

2

Metrics

3

Maximum difference

0

Recorded on 2 November 2026. Python 3.12.10, NumPy 2.4.1, Windows; all metric values are in percentage points.

02 / What the 216 rows mean

The script covers PolitiFact and GossipCop, 2/4/8/16-shot settings, seeds 1/2/42, and nine method or selection views. It reports Macro-F1, balanced accuracy and the proportion predicted as fake.

The 216 rows are output records, not 216 newly trained models. FCP-fixed and FCP-tuned may refer to the same saved run. The manuscript’s 344 archived training runs describe a different scope.

03 / Recompute it

Run these PowerShell commands in a new working directory. The two SHA-256 values matched in the recorded environment. Other platforms should compare parsed row identities and numeric values if formatting differs.

git clone https://github.com/GuHang2011/frozen-clip-fake-news-fusion.git
cd frozen-clip-fake-news-fusion
git checkout f06dd9d3b48ecd2bb4e3c6882d36de1ceb8cb9b5
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install numpy==2.4.1
$env:P2R = 'results'
.\.venv\Scripts\python.exe -I analysis/analyze3.py out
Get-FileHash analysis/per_run_operating_points.json -Algorithm SHA256
Get-FileHash out/per_run_operating_points.json -Algorithm SHA256

Expected SHA-256: da6c7575df0abb789325d461d41a77ecf881c247094ccc6dfc996305d05b56c8

04 / Scope and limitations

• This check confirms that analyze3.py recomputes the reference table from the supplied artifacts. Independent sample alignment and data provenance need additional evidence.
• Original training, bootstrap aggregate analyses and reconstruction of original data splits were not run.
• Hard predictions cannot recover ranking-based AUROC, probability calibration or a sweep over decision thresholds.
• The paper remains submitted to Neurocomputing, not accepted. This is a reproducibility check, not a new model result.