Contextual Contamination: The Silent Drift of Large Language Models via Stored Conversation Data

Thank you for pointing out this problem. I think the distinction is important: a model can remain locally coherent while still being influenced by the register, framing, or manipulative structure of the context it is supposed to analyze. I am working on DESi, a process architecture for evidence-aware LLM workflows. One part of it is context hygiene: separating raw adversarial or emotionally loaded material from the generation context, transforming it first into explicit claims, tactics, risk markers, constraints, and audit state. Would you be comfortable if I used your contextual contamination setup as inspiration for a small DESi benchmark? The goal would not be to claim ownership of your idea, but to test whether a DESi-style pipeline can reduce register drift, framing leakage, role adoption, and attribution loss compared with a baseline LLM that receives the raw context directly. If I run such an experiment, would you be interested in seeing the results?