AntonymGuard

Rejects matches where one prompt asks for the opposite of the other.

"How do I enable two factor authentication on GitHub?" and "How do I disable two factor authentication on GitHub?" differ in a single token out of nine. Every other signal a cache can read — length, structure, topic, word overlap — says these are the same question. They are the same question with the answer inverted.

The guard looks for a flip, not merely a difference: one prompt must use a word more than the other while the other uses its opposite more. Two things follow from counting rather than testing membership.

It stops false rejections. "Run this before deploy" and "run this prior to deploy" differ by the word before, but nothing anywhere says after, so there is no flip and no rejection.

And it still catches flips hidden by an incidental repeat. "How do I turn on format on save in VS Code?" and "How do I turn off format on save in VS Code?" both contain on — the second one in "format on save" — so asking whether on appears in both says they agree. Counting says the query uses on twice to the candidate's once while the candidate uses off and the query never does, which is exactly the swap that matters.