This audiobook is narrated by an AI Voice. Your system is working. Every test passes. The dashboard is green.

A model refused to answer while the answer sat in front of it, complete and correct, and then invented a data quality problem to justify the refusal, in the register of a careful bug report. A filter identified sensitive material perfectly, labelled it, wrote a log line proving it had seen it, and let it through. An upgrade rewrote the format of more than a quarter of responses while every accuracy test reported a clean migration. A cost model was wrong in two directions at once.

Six failures. Zero errors. Several of them producing positive evidence of health.

None of it was hidden. All of it sat in plain text, in fields nobody aggregated, phrased in a way nothing matched, or summarised by a statistic that absorbs exactly that signal. This book is about the class of failure that leaves no trace your monitoring was built to catch, and the five cheap numbers that catch it instead. None of them needs a labelled dataset, a vendor, a budget or anybody's permission.

The good news is the argument. The unpredictable part of your system was always the small part. The provable part was always most of it, and almost nobody tests it, because the presence of a model made an entire subsystem feel unmeasurable when it is not.

For engineers and technical leaders running something real, who suspect the green dashboard is not the whole story and want to know exactly where to look.