Answer overrides: when you need the exact words, every time
An override doesn't rank higher than retrieval — retrieval doesn't run at all. That's why it's guarded more tightly than an ordinary article.
Every other answer your assistant gives is assembled from retrieved passages. An override is the one exception: its text is delivered verbatim, under your organization's own authority, with no retrieval step in between. That's exactly why it's built to be harder to change than an article, not easier.
What an override actually does
Not "ranks first" — retrieval does not run. The moment a question matches an override, the search that would otherwise find and rank your articles never executes. The response is delivered exactly as written, never paraphrased, never blended with anything else.
This is why an override is not a general-purpose fix. Reach for one and you've told the assistant to stop reading your knowledge base entirely for that question — a required "risk reason" field forces you to write down why, in front of whoever reviews it.
Active and in force are not the same thing
An override can read Active in the system and still not fire, if its effective window has closed. The status badge computes the window and shows it as a separate fact, so a card that looks live isn't silently dead.
Every save lands in Draft — on create and on edit alike
Editing a live override un-approves it. It stops being delivered until someone activates it again, deliberately. Several fields are locked once an override exists at all: the match kind, the language, and (for an exact-question override) the question text itself. Change any of those and you need a new override, not an edited one — because a match rule that quietly changed shape would leave the old data pointed at the wrong logic.
The tester runs the real matching logic
Type a question into the tester and it runs the exact same normalization and matching rules production does — not a second, drifting approximation. If it says "no," it tells you why: the keyword didn't fire, or you're outside the effective window.
If an override cites an article that's since moved outside the list the editor loaded, the form refuses to silently drop the citation. It asks you to explicitly confirm the link is gone, rather than saving a null nobody meant to write.
Retiring is permanent
There's no re-enabling a retired override — retirement is the terminal state, with its own confirmation step, separate from the ordinary activate/deactivate cycle.