AnswerAnchorGuard

Rejects a candidate whose stored answer repeats the very word that the two prompts disagree on.

The near miss this exists for is the one the prompt-side chain is blind to by construction: two prompts that differ by a single substituted term the guards have no vocabulary for, so nothing fires, and similarity is high because the wording is otherwise identical. what is the capital gains tax rate when i sell a second home against …a primary residence clears every guard in MatchGuards.standard. The evidence that they need different answers is not in either prompt — it is in the cached answer, which opens "Gain on a second home is taxable in full".

So the rule is: when the answer names the term the query replaced, the answer was written for the other question.

What has to hold before it will reject

Three conditions, and each one is load-bearing:

1. The prompts differ only by a substitution. Same number of content words, differing in at most maxSubstitutions positions, compared with Text.isSameWord so a typo or a spelling variant is not a substitution. This is the same rule Text.differsOnlyBy applies to marker guards, for the same reason: a word only counts as evidence when everything around it matches. Without it, How do I configure CORS on an API gateway? against How do I configure cross origin resource sharing on an API gateway? is refused, because the answer to the second says "origin" — an expansion read as a swap, and a genuine paraphrase thrown away.

2. The substituted word appears in the response. Matching is Text.isSameWord again, so residence in the prompt is found as residences in the answer.

3. The query does not use that word anywhere. Which is better for session storage, Redis or Memcached? against the same two names swapped is a substitution at two positions, and the answer names both — but the query names both too, so the answer is not anchored to one of them. Dropping this condition refuses every word-order pair in the corpus.

What it deliberately does not do, and what it cannot do

It does not read numbers out of the answer. A response carries the figures that answer the question, and the question rarely carries them — what is the boiling point of ethanol names no number and is answered with 78.4 — so a numeric comparison against the response refuses honest paraphrases in bulk. ResponseGuardTest measures that rather than asserting it in prose. This guard reads only the words the two prompts have already singled out, which also keeps it linear in the response length with no parsing.

It also inherits Text.isSameWord's tolerance in both directions. mg against mcg is one insertion, so it reads as a typo and no substitution is seen — that near miss is not caught here. The same tolerance is what stops organise against organize being read as a swap, and this guard cannot have the second without the first.

It is opt-in, and the reason is the measurement

Of the 118 near-miss lookups that MatchGuards.standard still serves on the two blind corpora it refuses 14, and it refuses none of the 164 paraphrase lookups still being served. That moves the false-hit rate from 0.291 to 0.238 on the held-out split and from 0.333 to 0.309 on validation.

A guard that catches an eighth of the residual for no measured cost would normally go straight into the default set. This one does not, and the reason is the evidence rather than the result: those numbers are in-sample. The answers behind them were authored for this measurement, because no corpus of real paired answers exists to harvest, so it is a regression check rather than the blind measurement every other guard is held to. See docs/CORPUS.md and MatchGuards.responseAware.

Parameters