Reproducibility package for "Preprocessing-Induced Decision Instability in AI-Generated Image Detection"

Code, pre-specified analysis plan and complete study outputs for the manuscript “Preprocessing-Induced Decision Instability in AI-Generated Image Detection”. The study evaluates how spatial preprocessing (native cropping, resizing before cropping, and an equal-downsampling window policy) and nine encoding conditions change forensic features and individual decisions of AI-generated image detectors. It uses 300 RAISE-1k scene groups with DALL-E 3, Midjourney v5 and Firefly images from Synthbuster, five repeated grouped splits and leave-one-generator-out evaluation. Detectors: FALCON (nine entropy, JPEG recompression and Gaussian residual features with logistic regression), a CLIP ViT-L/14 linear probe, and UnivFD as a reference. The package contains the Google Colab notebook, Python modules, the analysis plan with its SHA-256 hash, every test decision and score, per-cell metrics, decision-stability and Rashomon tables, feature analyses, tables and figures. Source images are not included and must be obtained from RAISE and Synthbuster under their original terms.