How often we are right
Machine-measured fields carry a published accuracy figure, taken by hand-checking random samples against the books. The number is the warranty: it appears here, it ships inside every dataset, and it is what a licence is written against.
Unsampled fields stay out of the reader-facing filters and out of the licensed feed. That is the whole reason this page exists — the alternative is a confidence we have not earned.
| Field | Method | Checked | Confirmed | Shown to readers |
|---|---|---|---|---|
archetypes |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
setting_place |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
setting_period |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
world_type |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
species |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
pov_count |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
timeline |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
dynamic |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
door |
machine-extracted, evidence required | 0 | sample too small | no — below the 90% / 100-sample floor |
publisher, first_published | two lineages agreeing | n/a — Tier A is agreement, not sampling | yes ✓✓ | |
| every claim | editor judgement, signed | n/a — opinion, attributed, never scored | yes, with the reason | |
How a sample is taken
The desk is dealt ten random extracted facts at a time and confirms or rejects each against the book. Rejections do not just lower the number — they keep the field out of the reader filters until a later round earns it back.
The floor
A field needs 90% confirmed over at least 100 samples before it appears in Discover or ships in the default feed. Below that it stays in the pipeline, visible only here, where the failure is on the record too.
Tier A facts are not sampled because they are not guesses: they exist only where two independent sources state the same value. Claims are not scored at all — they are opinions with a name and a date attached, which is a different kind of accountability.