> arbitrary keys, decided at runtime" vs "fixed set of fields decided at definition time" (can't an NT's keys also be interpolated from runtime values?)

If you want to create a named tuple with arbitrary field names at runtime then you need to create a new named tuple type before you create the instance. That is possible, since Python is a very dynamic language, but it's not particularly efficient and, more importantly, just feels a bit wrong. And are you going to somehow cache the types and reuse them if the field names match? It's all a bit of a mess compared to just passing the entries to the frozendict type.