Data and methods
Versioned inputs and intermediate calculations make a study reproducible.
In plain terms
A study is only as good as what can be handed to someone else. Every number reported here comes with the inputs, the labels and the intermediate steps needed to recompute it.
Reported study scope
The preliminary comparison contains 20 true, 20 false, and 20 unsettled statements, with 239 usable replies from 240 requests. Original-label totals are 14/60 and 58/60. The study has no action examples; its totals are not estimates of action safety or deployment accuracy.
Study records
A reproducibility package includes exact item wording, label instructions and review history, raw model outputs, full prompts, routing, source assignments, confidences, vectors, normalization, parameters, and every intermediate decision quantity.
Evaluation discipline
Labels are established independently of scoring. Development and holdout data are separated by topic, origin, or time as appropriate. Calibration and thresholds are frozen before evaluation. False Allow, false Block, Review coverage, action errors, and uncertainty are reported together.
Research access
Use the contact page for research correspondence. Existing collaborator access remains available through the secure portal.