Failure mode

A technical term or concept is used as if the reader already knows it, with no inline definition or gloss. A link is not a definition — the reader should not have to follow a link to understand the sentence.

Test

On first encounter of each technical term, ask: does the surrounding sentence define it, paraphrase it, or give enough context to infer its meaning?

Exceptions — do not flag: - Standard technical vocabulary (LLM, context window, prompt, token, API). - Terms whose opacity is already covered by the notation-opacity gate (e.g., "external symbolic state" when the real access barrier is the K notation it labels). Do not double-flag the English phrase alongside a notation-opacity finding for the same concept.

Active vocabulary terms declared in AGENTS.md are NOT exempt. Authors may know these terms from loaded context, but external readers do not. On first meaningful mention in authored prose, provide both an inline gloss and a link to the definition note. The gloss lets the reader keep reading; the link lets them go deep.

Example (fail)

"An execution boundary usually creates two different questions"

Example (fail — active vocabulary without gloss)

"The skill is produced by distillation from the methodology notes."

Example (pass)

"An execution boundary — any point where one LLM call ends and another begins — creates two distinct decisions"

"The skill is produced by [distillation](./definitions/distillation.md) (directed context compression) from the methodology notes."