Retrieved, verified, then quoted. Never written.

Legislation is a controlled vocabulary, and nobody asks a question using it. “How high before I tie off” shares no words at all with “fall protection system”. So a question is expanded through a table of trade-to-statute bridges before it is matched — tie off reaches fall protection, anchorage and fall arrest; swing stage reaches suspended work platform.

Your own words are weighted more heavily than the ones we infer for you, and the bridges that fired are listed under “why this ranked here” on every result.

Keyword matching (BM25) catches the exact statutory vocabulary, which matters because users often quote the law back at us. Meaning-based matching catches the far more common case where the words differ entirely.

The two are merged by reciprocal rank fusion rather than by adding their scores together, because those scores are not on comparable scales and normalising one against the other would invent a precision neither has. When both methods pick the same provision, that agreement raises the confidence figure — it is a stronger signal than either alone.

A section whose only job is to define a word is ranked down unless the question is itself definitional. Definitions match keyword-dense queries extremely well and are almost never what someone asking “what do I have to do” wants.

Meaning-based matching is currently unavailable on this deployment, so results come from keyword matching alone. Rather than hide that, every answer subtracts 8 points from its confidence and names it in the breakdown. Questions phrased in trade language are easier to miss this way.

This is the part worth being precise about. A language model is handed the provisions retrieval already found and asked to return them in a better order. Its response is parsed as section identifiers and relevance numbers.

• Identifiers it returns that were not in the candidate set are discarded, so it cannot introduce a provision — and therefore cannot introduce text.
• Candidates it leaves out are appended in their original order, so it cannot bury a provision either.
• The response type has no field for a summary, an explanation or a recommendation. There is nowhere in the data structure for an interpretation to live.

A prompt is a request. The parser is the guarantee.

Before an answer is returned, each quoted passage is matched against the section text stored at ingestion. The only tolerance is typographic — curly quotes, dash widths, non-breaking spaces. Capitalisation is not normalised, because legal drafting distinguishes a defined term from an ordinary word by its capital letter.

If any single passage fails, the entire answer is withheld rather than partly served. That is deliberate: if the tool got one passage wrong, its handling of the others is not something you should be asked to trust on the spot.

A quote that drops a “not”, softens “must” to “should”, changes a threshold, or splices two real fragments into one continuous passage all fail this check. Those cases are covered by tests that run on every build.

Documents are fetched from official consolidation services, parsed per source, and then checked back against the source text. A document is withheld from answers entirely if it falls below 99.5% fidelity, or if the parser found fewer sections than the document contains.

The second check matters more than it sounds. A parser that finds a third of a regulation at perfect fidelity is more dangerous than one that finds all of it imperfectly, because the sections it missed reach you as silence — and silence reads as “there is no such rule”.

34 documents are indexed and 7,191 sections are searchable. Coverage lists every one, including the documents currently held back and why.