Conversation text hidden. Evaluation conversations are seeded with user prompts from OpenAssistant (oasst1) and ToxicChat (lmsys/toxicchat). ToxicChat contains real user prompts with toxic and sexually explicit requests, and model thinking and replies often restate or act on them, so user, thinking and assistant text is hidden by default and removed before it leaves the server. System prompts, role labels, colors and probe probabilities are unchanged. Researchers who need the full text can email saifullah.saifullah@dfki.de for an access code.