The current Meddies implementation preserves more provenance than the interface exposes. Knowledge search keeps a retrieved passage with its document ID, chunk ID, section heading, and page references when available. The response rules tell the model to cite dosage sources and evidence levels. A separate safety check can flag dose text with no citation-like marker.
But the clinician still cannot open a complete trail from a claim in the live answer back to the passage that supports it. The stored identifiers are useful raw material. The instructions and checks are useful controls. None of them, on their own, verify that a named source exists or says what the answer claims. We should not call this end-to-end verification yet.
A citation can still fail
A citation is a pointer to the source an answer claims to use. It fails obviously when the paper, protocol, or section does not exist. It fails more quietly when the source is real but irrelevant to the sentence beside it, or when a link opens an entire guideline without showing where the support is. Grounding is the next test: does the source content support the wording, scope, and conditions of the claim? Traceability goes further by giving the reviewer a usable path from that claim to the evidence. An answer can have a citation and fail both tests.
Even a complete source trail does not make the clinical decision. It exposes the basis so a clinician can decide whether the recommendation fits the patient. The World Health Organization's guidance on AI for health says humans should remain in control of medical decisions. The FDA's 2026 guidance on clinical decision support uses a health professional's ability to independently review the basis of a recommendation as one criterion for determining whether certain software functions qualify as Non-Device CDS. Neither says that a citation makes an answer safe. These are design references, not a claim that Meddies satisfies a particular regulatory framework.
The source trail has two destinations
For external evidence, the trail should end at a named document and the passage that supports the claim. For a statement about a patient, it should end at the relevant chart entry, including the record type and date. If the source is weak or two sources disagree, the answer should show that limit rather than turn it into a confident conclusion.
That still covers only what made it into the answer. Traceability cannot reveal an important fact that a chart summary omitted; omission risk needs a separate check.
The standard we are building toward
"Every answer shows its work" means more than placing a familiar reference beside generated text. The product has to connect each claim to the passage behind it, show which patient facts were used, and make missing evidence visible. When that path breaks, the answer should say so.
Meddies does not complete that path today. Until it is visible and testable, the presence of a citation shows only that an answer contains a reference. It does not show that the reference supports the claim.
