7/ The takeaway: if you only model the variables you built the task around, you may miss what steers behavior and what the circuit normally does
Sometimes "nuisance" variables *are* the behavior.
E.g. history-dependent "lapses" only look irrational in artificial worlds where nothing changes!