Rubric: cached answers

Written by a person. Last read by a person on 2026-09-06, 2 days ago. Its facts were checked by the eval suite on 2026-09-07.

You got an answer labeled cached and want to know what that means, or you are adding a page and the build has told you it needs a cached answer.

Enforced by check_cached_answers.

The claim

When the helper will not call a model, a reader still gets a useful answer for the page they are on, and it is labeled so they can tell.

Observable evidence

• The entries in data/cached-answers.yaml, and whether each names a page that publishes.
• Whether a default exists for pages with no entry.
• The provenance label on what the proxy returns.

The pass bar

Every entry names a page that exists. A default is present. Cached answers carry cached-demo, distinct from generated-live.

Why this rubric exists

These are not a degraded mode. They are what every reader receives today, because no key is configured, and they remain what a reader receives whenever a rate limit or the spend ceiling is reached. A fallback that apologizes is a fallback nobody wanted to build and everybody uses.

They are written by a person rather than captured from a model. A recorded generation is a snapshot of one prompt on one day, and this is documentation.

The label carries the weight. A reader who cannot tell a cached answer from a live one cannot judge either, and the difference matters most exactly when the answer is wrong.

Ambiguous cases

An answer keyed to a page that has moved fails. It would be served to nobody and reported by nothing, which is the quietest way for content to rot.

A page with no entry passes, and gets the default. Writing an answer for every page would produce thirty-five paragraphs of the same advice, and a default that says what it is beats a specific answer that is generic.

Whether an answer is any good is out of scope here. The check counts and resolves; whether the sentence helps a stuck reader is a fresh-eyes judgment.