D-A3-4: Citations are real, and that is free

Status: Accepted, implemented (data/fetch_education.py, Phase 1).

Context

A citation-grounded patient assistant is only as credible as its citations. Fabricating a plausible-looking source, or substituting a similar-but-wrong page when the exact one is missing, is not a data-quality problem in this project — it is a category error, because citation coverage is the design constraint that keeps the app outside the FDA's Clinical Decision Support device definition (docs/regulatory-basis.md, §6.1 in the build plan).

Decision

Education citations resolve to real, public-domain sources — MedlinePlus Connect and RxNav/RxNorm — free and key-less, never LLM-generated content. The LOINC code table itself is not a dependency at all: Synthea's own observations.csv already carries the codes, so the exact-match citation path never needs to license or ship the table (store/models.py::Observation.loinc_code). Where no vetted page exists for a code, tools/lookup_lab_education.py returns a declared CoverageGap — never a substituted near-match and never generated text.

tests/tools/test_tools.py::test_lookup_lab_education_declares_gap_for_uncovered_code and canonical case 14 (evals/canonical.py, "what did my {analyte} come back as?") assert the declared-gap path directly: the answer states the gap, it does not fabricate a result.

Consequences

• Every citation a reader sees is a link they can actually open and verify — the FDA CDS "independently review the basis" criterion, satisfied by construction rather than by prompt instruction.
• Coverage gaps are a measured build statistic (docs/evidence/data-profile.md), not a surprise found in production.
• NOTICE.md carries the exact attribution MedlinePlus Connect and RxNav require; nothing here copies a MedlinePlus page verbatim.