Evidence · Search · Abstention
Provenance, abstention and trustworthy engineering search
Useful engineering search must show where a claim came from, when it was observed, which scope authorized it and when no supported answer exists.
Written by WorldIR · Reviewed by WorldIR Engineering · Updated 26/08/2026
A search result is not evidence merely because it is relevant. Engineering work also needs to know which source made the claim, whether that source is permitted, which revision applies and how fresh the observation is.
WorldIR separates stable facts from observations. A standard dimension may remain valid across many snapshots. Supplier stock may expire within minutes. A capability may be public-source, self-asserted, document-verified, transaction-verified or independently audited. Those distinctions belong on the field, not behind one marketing badge for the whole organization.
Private scope is part of relevance
A technically relevant document from another tenant is a forbidden result. Public search can be combined with exactly one authorized tenant overlay. Cache keys, indexes, citations and timing behaviour must preserve the same boundary.
Abstention measures coverage honestly
An empty result is sometimes the correct result. WorldIR reports abstention and unknowns instead of filling them with approximate matches. A rising abstention rate is a coverage finding—often a signal to improve catalogue data—not a number to hide.
Knowledge claims require causal tests
More indexed data does not prove that retrieval improved an agent. WorldIR’s evaluation design freezes snapshots and compares the same task with no knowledge base, public knowledge and one authorized tenant overlay. Source families and future data are held out together so mirrors and renamed copies cannot leak answers across the boundary.