Use the same questions
For each candidate, record task fit, date range, eligible volume, rights status, connection quality, outcome evidence and missingness. Ask whether findings came from a selected sample or a full scan.
Check what the score conceals
A combined ranking can hide a critical weakness. One dataset may have excellent linkage but no clear outcome labels; another may have fewer records with strong verification. Keep dimensions visible and explain tradeoffs.
Evaluate within controls
Use authorized samples and a documented evaluation plan. Prevent contamination between training and held-out tests. Record limitations and avoid generalizing one benchmark result to unrelated production tasks. Quality evidence is an input to selection, not a substitute for evaluation.
Practical guidance, not legal advice or a buyer specification. Examples are illustrative. Reviewed October 20, 2026.