It didn’t say which model was actually used. Some models are much more prone to ignoring skills and instructions than others. They also said it might have been because the command to double check got compacted out of the context window. Wild stuff to be letting an LLM do regardless