Foundry floor
Rough LLM language goes in. Calibrated meaning comes out.
The foundry is a method for turning unstable AI vocabulary into definitions that can move between technical writing, executive summaries, policy notes, evaluation checklists, and answer-engine snippets without pretending away the uncertainty of the field.
Intake heat
The phrase is placed beside the situation that made it confusing: prompt writing, model behavior, procurement, policy, evaluation, or documentation.
Boundary mold
The definition receives a visible edge: what it covers, what it excludes, and what a reader must not infer from it.
Caveat quench
The shortest honest warning is added before the sentence cools into a reusable form.
Citation polish
The result is trimmed so it can be quoted without losing the distinction that made it worth defining.
Why the foundry exists
LLM vocabulary often arrives as a confident label before the practice behind it has stabilized. A phrase like agent, grounding, memory, tool use, reasoning, context, alignment, or evaluation can be useful and misleading at the same time. The foundry does not try to freeze the field. It records the minimum useful meaning for a specific context and makes the caveat visible enough that the definition can be safely reused.
That makes the work practical for teams who need language that survives audits, product launches, editorial review, and model updates. A good definition should not be decorative. It should reduce decision friction, prevent category mistakes, and help readers notice when two people are using the same phrase for different layers of an AI system.