Spanish AEPD corpus test

Watch out

The captured page is the announcement post, not the article, so the substance is one step out of reach. What is visible: the post says they tested on a legal corpus, that in short it works and is worthy, the article title Testing TypeSafe's Jev - Legal Corpus, and a truncated summary line reading We replaced a regex with a probability. It was right and we were wrong, describing a day spent on a Spanish legal corpus, not a demo, and promising what held up and what broke. Not documented anywhere in the capture: corpus size, which regulator or dataset the corpus comes from, how labels were derived, agreement or error figures, per-item cost, latency, which cases broke, and whether Spanish accuracy was compared against English. The option set, the prompt and how disagreement with the regex was adjudicated are equally absent. The thread is reactions rather than findings: one reply asks for examples, and the confident observation that it fits best as a pre-filter, where a confidence threshold sends only doubtful cases to the expensive model, is a third-party opinion with no measurement behind it. Treat the regex comparison as the author's conclusion on their own corpus, not as a general claim that rules lose to probabilities. Copy reviewed against the linked page on 2026-10-04.