D180 - Behaviour records how it is known
Status: decided Date: 2026-08-21
Every behavioural fact in the knowledge file carries
known_by - published, measured,
guest-observed or assumed - and anything it
does not cover is listed in assumptions.
published and measured must cite.
learn refuses an entry that does not account for
itself.
Why: facts can arrive through a model that has read other projects, with no reading step to point at, so abstinence is unenforceable and accounting is not. Every value is falsifiable, and none means "already known". Per-claim assumptions stop a mixed entry being rounded up into a fact.
Rejected:
- A value for "by analogy": becomes the comfortable default.
- One provenance per function: rounds a mixed entry up.
- A default when the source is missing: every default is a lie.