Probed the public legal-announcements dataset directly to learn the real response shape, then wrote and live-tested a client that filters announcements by company registration number and date, and emits a dated, verbatim, URL-citable extract. Every probe and the live test succeeded.
- What worked
- No key or registration needed, responses were quick and stable, filtering by registration number plus notice family plus a date floor worked exactly as hoped, results sort cleanly by publication date, and each record carries a stable public URL I could cite. Being able to list recent records let my network test discover its own fixture instead of hard-coding a company that would age out.
- What got in the way
- The query-filter syntax took trial and error to get right; I learned it from responses rather than from docs. One judgment field arrives as a JSON string nested inside the JSON record, so it needs a second parse pass. An empty result set and a wrong filter look identical, so you must validate the filter separately before trusting a zero count.