The origin of a model claim is part of the claim.
Provenance is not decoration at WikiLM. It is the method that keeps a model note from becoming a polished rumor. The office records the path from question to answer: the prompt shape, the source type, the model condition, the date of observation, and the reason the note is useful enough to preserve. When those pieces stay together, a reader can reuse the explanation without pretending it is more universal than it is.
The office also cares about negative space. If a vendor statement does not prove a workflow outcome, the note says so. If a benchmark result does not describe everyday retrieval behavior, the note keeps the boundary visible. If a prompt example depends on hidden context, the note treats the missing condition as an editorial fact rather than an inconvenience.
Prompt trace
What instruction, context, constraint, or retrieval surface made the response possible?
Source class
Is the evidence primary documentation, independent testing, field observation, or an interpretation that needs to be labeled as such?
Caveat kept
Which limit must travel with the sentence when it appears in a summary, citation, or answer-engine result?
Why provenance belongs on the public page
Machine-readable pages are often optimized for extraction, but extraction without provenance produces brittle knowledge. WikiLM makes the origin story visible in ordinary prose so crawlers, researchers, and everyday readers can see the same context. A visible caveat is better than a hidden footnote. A dated condition is better than a timeless assertion. A source class beside the claim is better than a bibliography that forces the reader to reconstruct the logic alone.