# Evidence rules

## Strength of signal, strongest first

1. **Observed behaviour** — they did it, you watched. Strongest.
2. **Behavioural trace** — analytics, logs, support tickets. Strong, but tells you
   what, never why.
3. **Recalled behaviour** — "I usually…". Moderate. People compress and rationalise.
4. **Stated preference** — "I'd like…". Weak. Label every one.
5. **Predicted behaviour** — "I would use…". Near worthless for prioritisation. Never
   base a recommendation on this alone.

## Counting

- Always `n of N`. Never a bare percentage on a sample under 30.
- Never "users said" for n=1. Write "one participant (P4)".
- Never "most" without the count beside it.
- If two researchers tagged differently, report the disagreement.

## Quoting

- Verbatim, in quotation marks, with a participant ID.
- Trim with ellipses; never paraphrase inside quotation marks.
- One quote per insight. A wall of quotes is a transcript, not a synthesis.

## Contradiction

When research contradicts the brief's assumption register, that is the highest-value
output in the study. Name the assumption number, state the contradiction, and say
what it costs to keep believing it.

## Honesty

State the sample, the recruiting method, and who was missing from it. A study of
five power users tells you nothing about first-run experience, and the write-up must
say that in its own voice rather than leaving the reader to infer it.