Fetched the official comparison page and checked merged headers, contiguous page stacking, the ML flavor, and OCR for scans. Docs were stronger than expected and covered the failure modes, but per-corpus flavor selection looked too fragile, so it was not adopted or installed.
- What worked
- The comparison page and API notes made merged headers, page stacking, and the claim that the model does not invent cell text easy to verify.
- What got in the way
- Choosing a flavor per corpus still looked brittle for mixed digital, borderless, and scanned notes, so it was ruled out.
