In such a case the right way to go is to build a formal model of the informal expectations first. I.e., split scientific and engineering phases. Collect your data, build a theory, validate it, and only then engineer an implementation of it, mechanically, with zero mental activity.
And what you seem to suggest, start writing a parser and then experiment with various real-world inputs until you're satisfied, is certainly not a very productive way of doing things.
I'm not suggesting anything but that you can't build it from first principles. Collecting data and building a theory is not working from first principles, it's just a more structured way of hacking it until it works. You're still liable to have the next website you crawl break your parser.
> but that you can't build it from first principles
Do not confuse principles (which are universal, simple and beautiful) with specs (which suck shit a metric ton per second).
> it's just a more structured way of hacking it until it works
You're confusing hacking with cargo cult coding. Hacking until it works is exactly this formal loop: collect the data, build a model, test if it's applicable. Cargo cult is "google for an answer, paste some code from stackoverflow, see if it works". The OP article is about the latter.
And what you seem to suggest, start writing a parser and then experiment with various real-world inputs until you're satisfied, is certainly not a very productive way of doing things.