Scientific assistants increasingly gather evidence through tools that aggregate multiple external sources. A record list alone does not reveal whether an empty result reflects a completed search, incomplete coverage, or total upstream failure. This paper develops a failure-transparency protocol that separates record identity, request coverage, and claim support. We combine three evidence streams: an audit of eighteen dated website examples containing twenty-four displayed records across eighteen source identifiers; an exhaustive analysis of 117 abstract source-state configurations; and a small live metadata-check episode. All displayed records carry a URL and source field, but the shortened examples do not expose a per-source execution envelope, illustrating why record provenance and coverage are distinct audit properties. In the controlled enumeration, record-only and HTTP-aware policies correctly classify 28 and 31 configurations respectively; a coverage-aware reference classifies all 117 by construction. These are contract checks, not language-model performance measurements. During three read-only Crossref requests, two returned HTTP 429 while one returned a retraction-notice record, providing a concrete case in which unavailable metadata must not be interpreted as a missing publication. Fifty-two offline connector checks also passed using previously recorded responses; this does not measure current upstream reliability. Illustrated reference, dataset, literature, and archive examples motivate a three-layer reporting contract and a practical answer-level evaluation rubric. The study offers a reproducible interface-audit method, not a new retrieval model or evidence of a commercial service's superiority.