READER: Robust Evidence-based Authorship Decoding via Extracted Representations

In three linesREADER is a provenance framework to identify source LLM from black-box responses without predefined prompts. Using activation mapping on a frozen proxy LLM and Bayesian evidence accumulation, it achieves 31-42% top-1 accuracy on single response and 70-84% on 50 responses on Agent500 (50-target dataset).

Summary generated by Claude — human-verified