Answers are grounded in this paper and cite their sources.
This record holds the preregistration as filed on OSF Registries, with its supplementary file. Before any new sample is generated and before any case that enters an estimate is scored, the registration fixes the rules for two evaluations of image-forensic detectors on western blot images. Direction A asks whether a detector trained to separate samples of one denoising diffusion probabilistic model from real blot crops ranks samples from one trained StyleGAN2-ADA model and one trained LoRA-fine-tuned latent diffusion model as well as it ranks samples from a fresh-seed retraining of the first model. The three new models are trained on the same 90% of the generator-training crops, and an in-domain detector per model is the reference for how hard each is to detect. Direction B measures how often two frozen duplication pipelines find documented band reuse within one figure and across figures of one paper, on three public benchmark sets (primary) and on cases documented in retraction notices, errata and expressions of concern (supporting), together with flag rates on material with no known manipulation. Two dated addenda, 3b and 3c, are deposited as separate records; they hold only the values the registered rules produce and are published before any new sample is scored. Only aggregates are reported, and no audited paper is named. Claude (Anthropic), an AI assistant, was used to help draft, evaluate and revise the plan.
No takes yet. Share an insight, caveat, or question.
Benjamin Haselhuhn (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: