Build and measure a hybrid PII detector

Measure before you trust: the gold spans come first, and the scorer counts a mention as found only if every character of it is covered. With that rule, each layer's contribution is a number. Patterns catch structured identifiers and, with one case-sensitive rule, upper-case proprietors. NER catches mixed-case names in prose. The LLM pass catches what neither can see: a first name alone, initials, health information, a NINO with an example prefix and OCR-garbled values. The decision that matters most is verifying every LLM span against the text, because the two errors it makes (an expanded name, a "corrected" postcode) would otherwise be redacted in a place where they don't exist, while the real value leaks.