How we will measure image-detection quality
This is a published test protocol, not a results report. A controlled holdout must be completed before we claim an accuracy rate for SUREvAI.
Collect and label the originals
Record real camera photos, human-made graphics, fully generated images, and AI-edited photos as distinct categories. Preserve creation records and license information. Include recent generators and images from multiple cameras and regions.
Hold out independent sources
Keep all versions of one source image together. Separate generator families, prompts, workflows, source photographs, and creator collections between development and final tests. Never tune thresholds on the final holdout or on a single memorable miss.
Create realistic transformed copies
For both real and generated originals, make JPEG, WebP, resized, cropped, and screenshot copies. Record the exact transformation so we can measure how delivery changes a verdict.
Report errors and time
Publish false-positive and false-negative rates, recall at a fixed false-positive target, Not Sure rate, and end-to-end completion time. Break these down by generator, image type, and transformation. Include sample counts and uncertainty intervals; do not turn a model score into an authorship probability without calibration.