forwalkin/VeriLens ? reverse-engineered prompt

Build me a reproducible benchmark that checks whether the image detectors from the original ai image authenticity scanner can really tell real photos from AI generated images.

I want the whole workflow to run end to end, from inspecting the Synthbuster source and pairing it with RAISE 1k, to building the dataset, auditing for obvious biases, creating the edited and AI edited variants, running a small validation benchmark first, then a full benchmark only if the sanity checks pass. Please include the report files and summary charts, and make sure the results show AUC with 95 percent confidence intervals overall and by generator, plus the original threshold based metrics like accuracy, precision, recall, F1, and confusion matrix.

Keep it reproducible, use the original detector behavior without changing the models, and make the final report help decide which detectors are worth porting to Android. If anything is unclear, look up the current docs online and make reasonable assumptions, but keep the process strict and deterministic.