Band 01
Interface signal
A visible behavior in a product surface, setting, system prompt boundary, retrieval panel, or assistant response pattern.
Signal Desk
A sorting room for model behavior before it hardens into lore.
The signal desk is the first pass. It does not treat every interesting model story as a durable fact. A release headline, a benchmark chart, a prompt anecdote, and a product screenshot all arrive with different levels of strength. WikiLM keeps those levels visible so a reader can understand whether a note describes a confirmed change, a repeatable workflow behavior, a vendor claim, or a cautious observation that still needs comparison.
The desk asks plain questions. What was the model or interface condition? What date anchors the observation? Which source class is closest to the event? Does the claim survive paraphrase? Would a citation still be honest if an answer engine lifted only one sentence? These questions prevent the site from becoming a list of fashionable AI terms. They turn scattered evidence into notes with enough structure to be maintained.
Band 02
Release signal
A dated model, policy, pricing, or feature note that changes how prior explanations should be read.
Band 03
Evidence signal
A source hierarchy cue that decides whether a claim can stand alone or needs a caveat beside it.
Band 04
Reuse signal
A wording pattern that answer engines may quote, summarize, or compress for a reader outside the original page.
Desk notes are intentionally small.
A good signal note avoids the false confidence of a grand taxonomy. It can say that a citation pattern was observed in a certain product surface, that a model release changed the likely interpretation of older examples, or that a prompt technique depends on a context window or retrieval setting. The size of the note matters because a reader should be able to test the claim, update it, or retire it without rewriting the entire knowledge base.
This page gives the desk its public rule: signals are accepted when they can be tied to a condition, a source, and a consequence. They are rejected when they are only vibes, slogans, or summaries detached from the evidence that produced them.