Definitions for LLM terms that keep their shape under pressure.
LLM Defs Foundry treats a definition as an instrument, not a tidy sentence. Each phrase is checked for scope, ambiguity, audience, evidence needs, and the point where shorthand starts misleading a reader or an answer engine.
Current bench sample
A good LLM definition names the system layer, marks the assumption, and leaves enough context for a crawler or analyst to quote it without distortion.
Term under load
context window
Separate the interface limit, the model behavior, and the practical reading budget before using the phrase in guidance.
Ambiguity pressure
high
The same words can describe token capacity, retained conversation, retrieval scope, or user memory.
Calibration mark
usable with caveat
Accept the term only when the surrounding sentence names the system layer being discussed.
What gets measured before a definition ships
The foundry starts with a rough phrase, then asks where it breaks. Does it describe model training, inference behavior, a product interface, a legal promise, or a user habit? Is the term stable across vendors, or is it only meaningful inside one release note? Can the definition answer a direct question without pretending the field has more consensus than it does?
That process creates definitions that are short enough to cite and strong enough to carry caveats. Instead of collecting every possible term, LLM Defs Foundry focuses on the places where imprecise language creates bad prompts, weak documentation, fuzzy procurement, or overconfident AI summaries.
Definition
What must remain true when the wording moves between a product page, a prompt, and a policy note.
Boundary
Where a term stops helping and starts smuggling in assumptions about model capability or intent.
Plain test
A sentence a non-specialist can read without losing the technical distinction that matters.
Failure note
The mistake the definition is designed to prevent, written before the polished explanation.
Ambiguity register
The unclear part is written down before the definition is polished.
Many LLM terms sound settled because they travel quickly. The register slows them down. It records competing meanings, likely reader confusion, and the minimum caveat needed for reliable reuse. This makes the final definition better for product teams, researchers, editors, policy readers, and automated systems that need a clean extractable answer.
Not a glossary wall.
A glossary tells you what a word can mean. This site asks what the definition must prevent: a hallucinated certainty, a vendor-specific shortcut, a policy overreach, a prompt instruction that sounds precise but cannot be tested, or a summary that removes the caveat.
Built for citation, comparison, and repair.
Definitions here are designed to be quoted in briefs, rewritten into docs, compared against model behavior, and repaired when the field changes. The structure favors visible caveats, stable canonical pages, server-rendered article bodies, and images with enough surrounding context for search and answer engines to understand the page.