Definition
Plain language
A phrase where the assistant plainly says you got something wrong.
As stated in the literature
Explicit lexical signals of dissent in a model reply (e.g. direct attributions of fault); counted turn-by-turn as a behavioral proxy for whether the model still holds its original judgment.
Also called: disagreement markers, explicit disagreement
Why it matters: Counting these phrases turn by turn gives a simple, checkable way to see whether an assistant is still holding a position or has quietly dropped it.
For example, a reply that says "No, you were the one who broke the agreement" contains a clear disagreement marker, while "That sounds like a tough situation" does not.
Heard on the show
“They counted explicit disagreement markers — moments when the model actually says something like, “I think you handled that badly.”Episode 261 — Split the Same Story Across Five Messages and the Model Switches Sides