When you're 6 procedure calls deep in a model you started from some abstraction on the problem which has hit "well ACKchewally.." variances from the model, it's high cost to go back to top and invent the top level abstraction again, to be only 5 levels deep when you hit the next one.

Putting a late stage "yea I know, but this one time..." hack in, is logistically simpler.

TL;DR we're not addicted to complexity, we're addicted to the abstraction we started with, even when it turns out not to be as good as we thought.

This rings almost true for me - but “addiction” makes it seem like some sort of personal failing, whereas I think the cost of that backtracking and the probability of finding another Balrog in the woodpile later makes it seem like a perfectly reasonable choice.

Source: been there, done that.