Assuring Agent Safety Evaluations By Analysing Transcripts — AI Alignment Forum