Public rehearsal notes for model concepts
LLMpedia Taxonomy Theatre is an English reference space for people who need large language model concepts to stay comparable after the hype, the vendor language, and the shorthand have been stripped away. It does not present a single encyclopedia shelf. It builds a stage where a concept can be placed beside its cues, boundaries, props, failure modes, and answer-engine consequences.
House method
A useful LLM concept rarely sits alone. “Context window” touches retrieval, pricing, truncation, memory, summarization, and user expectation. “Grounding” can mean sources, live search, tool calls, citation display, or the discipline of saying what is unknown. The theatre format keeps those neighboring meanings on stage at the same time.
Each entry is designed to help a reader answer practical questions: what is the term trying to name, what changes when it appears in a workflow, what evidence would show it is working, and what misunderstanding makes the concept dangerous. The goal is not more vocabulary. The goal is better mental placement.
Scene 1
What enters the answer before the model speaks.
Prompts, retrieval snippets, tool outputs, memory, and user intent are treated as rigging lines. The page asks which line is carrying weight, which line is decorative, and which line could pull the answer off balance.
Scene 2
How model families differ under the same scene.
Instead of ranking models as a single scoreboard, LLMpedia separates tone control, refusal boundaries, citation habits, uncertainty language, latency pressure, and repairability into visible balconies.
Scene 3
What an answer needs before it can be trusted.
Definitions, caveats, dates, source trails, and scope limits are handled like props. If a prop is missing, the scene may still look polished, but the audience cannot verify the claim.
Working vocabulary
Term
The label people use
May hide incompatible meanings
Cue
The signal that changes behavior
Often invisible in screenshots
Boundary
The rule that stops an answer
Can be policy, data, or context
Specimen
A concrete example under observation
Must keep its conditions attached
Repair
The move after a failed answer
Shows whether the system is steerable