Evidence methodology
How claims enter the evidence base, how they are evaluated, and what the publication policies mean.
Where the claims come from
The evidence base contains canonical claims about how search and generative engines behave, extracted from primary sources: official search-engine documentation and announcements, patents, peer-reviewed and industry studies, practitioner surveys and reputable industry analysis. Each claim is normalized into one precise, testable sentence, and keeps its provenance — which source said it, in which version, retrieved when — plus a relationship graph connecting claims that repeat, narrow, broaden or contradict each other.
How the CORE claims were evaluated
The highest-priority claims (the CORE set of 300) each went through two independent primary evaluation runs: an evaluator researches the claim against current, retrievable evidence and returns a verdict with the documents and passages that support it. The two runs are aggregated blind — neither evaluator sees the other. Agreement closes the claim; disagreement escalates to a third independent evaluator, and the hardest cases go to an evidence-weighted adjudication that reads only the evidence, not the earlier verdicts. The result is a final consensus verdict per claim, frozen as a release.
What the verdicts allow
Each evaluated claim carries a publication policy, and Editory’s documentation obeys it mechanically. Supported claims may be asserted as written. Contradicted claims may be repeated only to correct them. Historically-true claims require explicit historical framing. Conditional and overgeneralized claims must keep their condition or narrowed scope. Claims with insufficient evidence may only be discussed as uncertain. And every claim without a completed final evaluation stays “under evaluation” — plausible is not the same as established, and nothing is promoted silently.
Frozen releases
The evidence base ships as an immutable release: a single database artifact whose contents are pinned by a cryptographic hash and identified by a release and freeze ID — the ones shown at the bottom of every documentation page. The platform opens it read-only and refuses to serve evidence from a file that does not match. New evidence work produces a new release; nothing is edited in place. That is what makes an answer reproducible: the same question against the same release gives the same evidence.
What this means for you
When an Editory documentation page — or, in a later phase, an API or agent answer — makes a recommendation, you can trace it: recommendation → docs topic → claim ID → sources → release. If we cannot trace it, we do not assert it. Where the honest answer is “not established”, the documentation says exactly that; we consider that a feature, not a gap.