Vocabulary

Words that flip an answer, read by AntonymGuard. Symmetric: listing enable to disable also covers the reverse.

Cues that make argument order significant, read by DirectionGuard: comparisons and conversions. "A vs B" and "B vs A" are different questions; "sort a list in Python" and "in Python, sort a list" are not.

Negation markers, read by NegationGuard. Contractions are caught separately by an n't substring test, so isn't needs no entry.

Words capitalized by grammar rather than by reference, read by Text.entityTokens — the ones that would otherwise look like named entities wherever they appear.

Words that describe the shape of the answer rather than its subject, read by ScopeGuard: output format, length and depth.

Words that may open a sentence without naming anything, read by Text.entityTokens.

Function words, question words, politeness fillers and the fragments left behind when a tokenizer splits contractions (don't → don, t).

Absolute time references, read by TemporalGuard.

Unit and currency tokens, read by UnitGuard, each mapped to a canonical name.