Field method
Keep the setup visible
A model observation is only useful when the prompt, constraint, context, and correction path are visible enough for another reader to inspect the claim.
Read the methodLiving field guide for model behavior
Observe the model before you name the concept.
WikiLM Field Atlas treats language models less like static reference entries and more like changing field sites. Each note starts with an encounter: a prompt, a response, a correction, a refusal, a sudden improvement, or a quiet failure. The work is to describe what happened without pretending the behavior is simpler than it is.
PromptTraceRepair
Prompt weather
How the same instruction changes under role, context length, examples, and hidden assumptions.
Answer posture
Whether a model explains, hedges, narrates, refuses, compresses, or invents structure under pressure.
Trace residue
The remaining marks after a response: missing citations, overconfident bridges, repeated frames, or useful caveats.
Repair behavior
How a model responds when the observer corrects, narrows, contradicts, or asks for uncertainty.
What this atlas keeps
Not a glossary. Not a terminal. Not a database. A place for behavioral evidence.
Many AI knowledge sites begin by naming things: token, embedding, context window, agent, benchmark. WikiLM begins one step earlier. It asks what a person can actually see when a language model is asked to summarize a source, hold a constraint, explain a disagreement, or admit that it cannot know. That makes the site an atlas rather than an encyclopedia. The entries are arranged around observable terrain: ridges of confidence, pockets of ambiguity, brittle instruction paths, and response habits that only appear after a second or third turn.
The `.xyz` part of the name is intentional. This is a workshop for experimental notes, not a polished doctrine. A good field note can be small: one prompt pair, one comparison table, one failed instruction, one method for checking whether an answer carries evidence or merely sounds complete. The goal is useful orientation for researchers, editors, builders, students, and curious readers who need language-model behavior described in plain English without losing the complexity.
Behavior map
Read response terrain
The atlas treats confidence, refusal, compression, citation habits, and repair moves as terrain that can be compared across situations.
Open the atlas
Notebook stance
Prefer slow naming
A behavior label should arrive after evidence. WikiLM keeps notes provisional, source-aware, and readable enough for citation by humans and answer engines.
Learn the origin