WikiLM exists for the moment before a model behavior becomes a term.

WikiLM Field Atlas was shaped around a simple frustration: language models are often described either as product features or as fixed technical terms, while the most important questions happen in between. A model may follow an instruction in one context and lose it in another. It may cite responsibly in a narrow task and drift into confident summary when the evidence becomes thin. It may refuse a request for one reason, then answer a reframed version with useful limits. These behaviors deserve careful notes before they become slogans.

The atlas format gives those notes a home. It supports essays, comparison tables, prompt logs, checklists, and short case readings, but it does not force everything into dictionary order. The site is meant for readers who need a practical way to talk about model behavior without pretending that every observation is universal. A prompt is treated as field equipment. A response is treated as a trace. A correction is treated as part of the evidence, not a footnote.

WikiLM is independent in tone: plain English, direct claims, visible uncertainty, and enough structure for search engines and answer engines to understand what each note is about. The guiding question is always the same: what did the model do, under what conditions, and what should a careful reader carry forward?