Afterword

This book will be wrong in places, and the interesting question is which places.

The measurements will not be. They are computed by running the tools while the page is built, so a number here cannot drift away from the tool that produced it. That was the whole reason for building the book this way.

What can be wrong is everything around them. A number can be correct and its interpretation mistaken; an axis can separate two files for a reason nobody has noticed yet; a threshold fitted to a gap of 8% against 23% can turn out to sit in the wrong place the first time a document lands between them. Several of the constants in the fold are exactly that shape, and each one is written down with the gap it was fitted to, so that the day a document falls in the middle it is obvious what happened.

The largest open question is still the first one. Whether there is a capturable pattern under JSON work at all is not settled, and the honest state of it is that the evidence is encouraging and incomplete. The fold, the partition, the naming of keys that are data and the pricing of candidate rows do transfer across documents that have nothing else in common. That is more than was expected. It is not yet enough to build on.

The strongest argument against the whole idea is in the corpus too, and it is kept there deliberately: one document that defeats the central operation is handled by an existing library in a single expression. A project that hides its best counterexample is not measuring anything.

If it turns out there is no pattern, that result will be written up with the same care as a success, and no package will be built. That is not a consolation prize. It is the answer to the question that was actually asked.