It can only tell you what it already says

An assistant that composes answers about a scholarly record has one failure mode that matters: it produces a fluent sentence that nobody wrote and nobody checked, attaches it to a real object, and it gets quoted. On a record whose whole argument is that every claim carries its evidence, that would be self-defeating.

So this does not compose. Every sentence it returns already exists in the record, with the same confidence mark and the same source it carries everywhere else on the site. The ranking is a plain term-and-phrase match, computed in your browser — nothing you type leaves this page, and there is no model behind it.

The cost is that it will not answer a question phrased in an unexpected way, and it cannot join two facts together for you. That seems the right trade for a record that is asking to be trusted.

What it will never do

• Attribute, authenticate, or offer an opinion on whether the sheet is what it is said to be.
• Say anything about market value.
• Infer a watermark, a state or a date from anything.
• Fill a gap in the provenance with a plausible guess.
• Write anything into the canonical record.