Released in April 2025, MRCR plants several near-duplicate messages in a long context window and asks the model to identify a specific one by its position or content, rather than merely proving it can echo back an arbitrary needle. Needle position within the context and the generation seed both change the result, so both must be reported alongside the score.