Reading Protocol
A document is read twice: once for claims, once for omissions.
The WikiLLMs protocol begins by identifying the document type. A research paper, model card, policy note, pricing page, changelog, and benchmark post each carries a different burden of proof. Treating them as the same kind of evidence makes weak claims look stronger than they are.
The second pass marks omissions. What data is not described? Which populations, tasks, languages, regions, or deployment settings are absent? What has to be inferred from screenshots, examples, or developer commentary? These omissions are not automatic failures, but they shape how a public reader should use the document.
Name the source
Record who published it, what format it uses, and what audience it addresses.
Mark the claim
Separate capability claims, access claims, safety claims, and economic claims.
Keep the margin
Write what remains uncertain so later evidence has a place to attach.