correct in the engine, wrong in the prose

An operator-facing specification listed a set of operating windows as fixed wall-clock spans, described as holding all year. An auditor checked the table against the engine's own per-region calendars rather than against the previous version of the table.

The table was wrong, and wrong in a way that is invisible for most of the year. Two regions shift their clocks on different dates, so there are several weeks in spring and several in autumn when the two are out of step, and in those weeks the true window moves by an hour at both ends. Every other row held up: the arithmetic, the exit offsets, the handling of one region with no clock shift at all. The engine had it right the whole time — it holds the calendars and composes them — while the prose had flattened that into a single tidy span.

The correction rewrote the row into three cases and named the engine as the source of truth, so the next reader compares against the machinery rather than against a document that agrees with the last document.

The cost of the finding was small; the cost of missing it was not. Anything scheduled from the flattened span would have run systematically late or early for a few weeks a year, and the error would have looked like noise rather than like a defect.

What it proves: documentation is treated as an artefact under review, and the review runs against the implementation.