Learning by fixing other people's mistakes
A broken set of books is a teaching problem somebody has already solved for you, at their own expense.
A large share of the work I have been handed arrived already broken. Books that had drifted, a client base collecting late payment penalties, a newly acquired property portfolio with costs nobody had looked at, reporting nobody could explain the origin of. In each case the job was to make it correct, and in each case somebody else had made it incorrect first.
That is normally described as unglamorous work, and the description sticks because it is true about how it feels. What nobody told me is that it is the most efficient way to learn a domain I have ever encountered, and I only worked out why after I had stopped doing so much of it.
A broken thing is a labelled example
Study a correct system and everything in it looks equally necessary. You cannot tell which parts are load bearing and which are habit, because nothing is under strain. It is a photograph of a finished house. You can admire it and you learn almost nothing about which walls hold it up.
A broken one has already run the experiment. Something failed, visibly, and somebody is paying for it right now, which means the failure has a size and a date. And if you dig you can usually find the cause, because whatever went wrong went wrong at a particular moment for a particular reason and left a trace.
That is a labelled example. Input, failure, cause, all in the same artefact. In any other field you would pay for a dataset like that.
It also gives you the counterfactual, which is the part I value most now. In a working system you never find out what would have happened without the control, because the control worked. In a broken one you are looking directly at what happens without it. I know why certain small disciplines matter, not because anyone convinced me, but because I have spent weeks of my life reconstructing what happens in their absence.
Which makes the economics odd
Cleanup is priced low and handed to whoever is new. Two things fall out of that.
The first is that the richest teaching material in the field is systematically routed to the people least equipped to read it. A broken file only teaches you if you can already see what it should have been. If you cannot, it is just a mess, and you learn to survive it rather than to read it, which is a completely different skill and a much less useful one.
The second is that nobody is left to notice the patterns. The people who have seen enough broken systems to generalise across them have usually been promoted away from the work by then. The generalisation is the valuable part and it lives in exactly the people who no longer touch the artefacts.
What it does not teach
I overrated all of this for a while, and there is a real limit.
Fixing teaches failure modes. It does not teach design. A catalogue of ways things go wrong is not a picture of what good looks like, and I noticed the gap when I started building things from nothing rather than repairing them. I was quick at spotting what would break and slow at proposing a shape, because I had years of examples of the first and very few of the second.
There is also a selection problem in it. You only ever see the systems that broke loudly enough for somebody to hire help. The ones that were quietly mediocre for a decade never reach you, and those are probably the more common case.
The last thing is smaller and harder to shake. Cleanup work ends with the artefact in the state it was supposed to be in all along. There is nothing to show. The people who benefit cannot see what changed, because what changed is that a thing which was wrong is now ordinary, and ordinary does not photograph.