# Confidence

Confidence is ResponseHub's assessment of how good an answer is, given the
material it had to work with. It is the best single indicator of how much of your
attention an answer needs.

| Confidence | |
|---|---|
| **Excellent** | 80–100% |
| **Very good** | 60–79% |
| **Good** | 40–59% |
| **Ok** | 20–39% |
| **Low** | Below 20% |

Expand the rating on an item to see the reasoning behind it.

## How it is worked out

Confidence is not produced by the same process that writes the answer. It comes
from a separate, adversarial check that runs afterwards.

Once an answer exists, ResponseHub passes three things to a language model acting
as a judge — the original question, the answer that was produced, and the
[references](/how-questions-are-answered/references/) behind it — and asks how
well the answer is supported by that evidence. The rating and its rationale come
back from that assessment.

Judging the answer separately from writing it is what makes the rating worth
something. It is an independent second opinion on work already done, rather than
the same process marking its own homework.

## Using confidence to triage

The point of the rating is to tell you where to spend your review time.

**Excellent** answers tend to be highly reliable. A quick read is usually enough.

**Very good** answers are, more often than not, also fine.

**Good, Ok, and Low** answers need proper human review. Read them against the
references, and expect to edit.

**Approval is still yours:** Confidence never approves anything. Every answer needs a person to approve it
before it goes out, whatever the rating — see
[answering questions](/questionnaires/answering-questions/).

## When confidence is persistently low

Low confidence on a question usually means your material doesn't cover it well.
The fix is to get the answer into your knowledge base, so the next time it is
asked there is a curated answer to draw on rather than policy text to piece
together.

You don't have to do that at the time. Answer the question, approve it, and
finish the questionnaire — then add it to your knowledge base as part of
[updating the knowledge base after a questionnaire](/questionnaires/updating-kb-after-questionnaire/),
which offers exactly these questions: the ones your knowledge base could not
already answer.

## Next steps

- [Relevancy](/how-questions-are-answered/relevancy/) — and why confidence is the better indicator
- [Updating the knowledge base after a questionnaire](/questionnaires/updating-kb-after-questionnaire/)