Robin Saige we check the tools AI agents call

Can it be checked?

Before asking whether a server is right, the earlier question: could anyone even check? Every graded server sits on a ladder from take its word to answer key in hand. Coverage first: 5,807 of 6,040 answering servers graded.

the ladder applies to public-fact retrieval only

5,126 of 11,094 — 46.2% · actions, generators, and private data are checked differently — see method: by kind

The ladder

answer key in hand V3W — a public authority publishes the true answer AND we re-check this server against it today
13
has an answer key V3 — a public authority exists for its subject; wiring it is work, not impossibility
2,299
has a twin V2 — no single authority, but an independent instrument measures the same reality — agreement proves consistency, never truth
105
shows its work V1 — names a source or method, but no public copy exists to compare answers against
1,151
take its word V0 — states no source, no method — nothing anyone could check
2,239

rank order · as of 2026-08-14

Disclosure — does a tool say what it covers?

Three yes/no reads of every stored description: names a source · states coverage · states freshness.

0/3 disclosed3,196 1/3 disclosed951 2/3 disclosed1,544 3/3 disclosed116

as of 2026-08-14

Operator? How a server climbs →