This assessment is produced automatically. Four independent signals are read — user reviews, SDKs in the app package, the app's own on-screen text and the app manifest — and each candidate finding is filed under the sources that support it.

No charge stands on one kind of evidence — with one published exception. A finding is either a measured fact that anyone can reproduce (an ad SDK is in the package; every subscription bills weekly; the privacy label omits a tracker we can see), or it is an interpretation — and an interpretation needs independent sources to agree before it reaches the catch list. The exception is a harm only users can see: a charge may publish from reviews alone, as “Reported by N users”, once enough reviewers raise it to clear that charge’s own published bar. That agreement rule is what replaced the human reviewer, not a model’s confidence.

Findings under One source only are real reports we do not stand behind as findings. They are shown because withholding them entirely would hide a genuine complaint, and labelled because publishing them as charges would be an accusation off a single source.