# When organization NPI counts overstate address counts

For each of the 13 organization CSVs, count unique NPI rows and distinct nonempty location_key values. location_key is the existing normalized primary street line plus five-digit ZIP. Subtract distinct primary keys from NPI rows; divide that difference by NPI rows for the percentage. Count an address collision when more than one NPI in the same legal-name profile has that key. Do not merge records across legal names or include secondary addresses in this particular comparison.

A shared address is a matching candidate, not proof of a duplicate registration, a shared owner or a single operating site. The key does not independently verify the address, resolve suites or establish whether two departments have separate buying authority. The cohort is the 13 profiles already published by Provyx, selected rather than randomly sampled. Totals are sums of within-organization counts, not globally unique facilities.

Reproduce: python3 scripts/build_research_drafts.py organization-npis-vs-practice-addresses; then python3 scripts/build.py. Inputs and SHA256 hashes are in provenance.json.