Summarisation models that do not follow instructions already exist! Fixing exfiltration is good, low hanging fruit. But for a summarisation task, whole classes of typical instruction following behaviour are totally off target!