How we test, and what we have not proven
This library explains how SUREvAI evaluates its detectors. We publish protocols and source-backed notes while assembling a labeled test set. We have not published a SUREvAI-wide accuracy percentage.
Why one accuracy number is inadequate
A detector may work well on clean images from familiar generators and poorly on screenshots, graphics, or new generators. We plan to report results by source and transformation, including false alarms on genuine photos.
Current research notes
Our benchmark protocol defines how samples, labels, splits, and metrics are selected. Our compression note describes test conditions for TinyPNG-like recompression, JPEG, resizing, and screenshots; it does not claim a measured SUREvAI result.
When results are ready
Each future report should state the sample count, collection dates, detector version, thresholds, licensing, confidence intervals, and the full set of failures. We will distinguish our measurements from numbers reported by other researchers.