There is a quiet problem with many AI tools that cite sources: every citation looks equivalent. A reference marker can point to a practice guideline, a pooled analysis, a single observational study, or an unreviewed preprint. Retrieval has happened, but the user is still left to reconstruct the source context.
CliniAtlas therefore attaches an operational evidence stratum to each retrieved record and exposes it in the reference interface. The label is a navigation aid, not a complete critical appraisal.
The strata model
The current model separates guidelines, regulatory records, systematic reviews, randomized trials, observational studies, case reports, expert opinion, and preprints. This helps readers see what kind of evidence supports an answer and where more scrutiny may be needed.
Source type is not the same as certainty. A weak guideline can be less trustworthy than a well-conducted trial, and a systematic review can only be as reliable as its included studies.
For that reason, CliniAtlas does not claim that the strata replace GRADE, risk-of-bias assessment, applicability review, or clinician judgement. They are a transparent first layer that makes the retrieved set easier to inspect.
Why regulatory records are separate
EMA product information and assessment reports answer jurisdiction-specific questions that journal articles may not answer directly, such as authorized indications, approved conditions of use, contraindications, and regulator-reviewed safety information. They deserve an explicit category rather than being mixed into a generic literature list.
Sometimes the right result is a limitation
Labeling sources also supports calibrated answers. If a search returns only small observational studies or preliminary records, the interface should make that visible and the synthesis should state the limitation. Citation count alone is not evidence strength.
Measured before marketed
CliniAtlas has an internal evaluation harness for retrieval, citation faithfulness, and answer-level checks. Those internal tests are useful for engineering, but they are not a substitute for external clinical validation. A public methodology page and clinician-reviewed evaluation set are planned; until those are published, we will not present internal scores as independent proof.