This audiobook is narrated by an AI Voice. Six times during this programme, something the author had built reported a confident falsehood. Not one of the six was caught by a check he had written.

One of them was the strongest finding in the entire project: a perfect result, every trial, in the direction the whole field expects. It was completely false, and the only thing that caught it was opening a file and reading what the model had actually said.

This is the instrument book. How to build a comparison that means something. How to find out what a wrong answer scores on your own tests, before you report a score at all. How to prove a grader can fail before you let it grade anything. Why an evaluator's error rate tells you almost nothing and its error direction tells you everything.

Every method in it is a few hundred lines of ordinary code with no dependencies, and every one of them exists because something went wrong first.

It is also, unusually, a book that documents its own failures at full length: the detector that looked for one set form of words and reported ninety-six fabrications where there were none, the check that compared a string against a pair of values and therefore could never fail, the average built from six outliers of ninety-six. Not as a confession, but because the pattern in how they were found is the most transferable thing here.

For anybody who has to decide something about an AI system and would rather decide it on evidence than on somebody's confidence.