Answer model

Different models can compress context, phrase claims, and display links differently. Mixing their results makes citation changes uninterpretable. Record a name when the interface reports one, and do not infer a version from writing style.

Document retrieval and answer generation are different stages. A model may receive context from search, memory, a user file, or multiple sources. A model naming a page therefore does not prove it found or read that page during the current query.

Save the product, mode, displayed model name, settings, query, input documents, answer, and time. Mark hidden parts as unknown. Honest uncertainty is more useful than false precision and lets you repeat the experiment after a product update.

Use a control set with the same sources and run it in two modes. Compare facts, links, caveats, and omissions, not just wording. Repeat on another day: stability is a separate observation, not a property implied by a model name.

A public interface may hide routing and update a model without changing its displayed name. In public research, answer model is therefore often a label for an observed mode, not a proven description of the internal system. When a provider hides routing, the same model label does not guarantee the same answer model across dates.