An answer is only useful if you can inspect its source

Keep editions distinct

The service accepts voice or text questions against a rulebook catalog. Its source model records PDF provenance and checksums, preserves extracted line maps, and allows an active edition to be selected. Different copies are kept separate rather than silently merged.

The README documents actual discrepancies between supplied rulebooks and a cross-reference mismatch. That makes version handling a first-class feature: finding plausible text is not enough if the answer comes from the wrong edition.

Offer a useful fallback

Without a configured AI key, the application starts in clearly labeled source-search mode. With provider access, it can transcribe, navigate an index, compose a cited answer, and check that answer against evidence. PDF.js handles newly uploaded PDF extraction in the hosted implementation.

Long requests return account-owned job identifiers. A restart marks interrupted work instead of silently issuing another paid request. Question and result lifetimes are bounded.

Published PDFs remain immutable objects. Newly extracted line breaks can differ from earlier extraction tools, so archived citation maps are reused rather than casually regenerated.

Keep evidence within reach

The reader can inspect the source behind an answer. Automated checks reduce some errors but cannot establish every interpretation, especially when editions disagree. The catalog, immutable line maps, and source viewer keep that uncertainty visible and give a human reader a practical route from an answer back to the relevant rule.