AINPI·ainpi.dev
NDH·2026-07-13
Release update · 2026-07-13

No June release.The missing endpoints are already public.

Days since last NDH release
66
CEHRT endpoints in the public scrape
31,255
NDH orgs with zero endpoints (April)
98.7%

The NDH manifest still serves the 2026-05-07 files, 66 days on. Meanwhile the CMS directory team's public scrape of certified-EHR HTI-1 bundles holds 31,255 FHIR endpoints with NPIs and states attached, while 98.7% of NDH organizations carried zero endpoints in the April release. H45, pre-registered today, measures that gap state by state.

Read the full report →
methodology v0.7.2-draftNDH 2026-07-13

2026-07-13 update: no June release, and the missing endpoints are already public

Three things this update: the release cadence broke, the CMS directory team's public repos show exactly where the directory is heading next, and we pre-registered a finding (H45) that measures the gap they are about to close, state by state, before the ingest lands.

No June release

The NDH bulk-export manifest at directory.cms.gov still points at the 2026-05-07 file set. April came on the 9th, May on the 8th, and then nothing: the May release has now been the current release for 66 days. AINPI checks the manifest rather than re-downloading files, so this is a statement about what CMS publishes, not about our polling. Every AINPI number remains pinned to the 2026-05-08 release until a new one drops.

What the CMS directory team shipped instead

The directory's public development happens in the open, and the last two weeks were busy:

  • A scrape of certified-EHR FHIR endpoints. Under the HTI-1 rule, certified EHR vendors must publish service-base-URL bundles: organization name, address, often an NPI, and the FHIR endpoint URL. ONC's Lantern aggregates them, and the CMS team's npd_ehr_fhir_npi_slurp pipeline now flattens them into joinable records. The public cache holds 7,999 Organization entries and 31,255 Endpoint entries across 17 vendor hosts (2026-07-10 snapshot).
  • A payer-data downloader (npd_slurp_payer), the second-generation pipeline for pulling public payer directory data into the NPD.
  • A Federated Payer Identifier proposal: payers self-enumerate, publish a well-known index file mapping plans to endpoints, and the index files are mirrored in git so corrections arrive as pull requests. Early prototype, explicitly not authoritative, but it is the first concrete mechanism we have seen for the payer half of a National Provider and Payer Directory.
  • An updated submission spec for provider affiliations: a PractitionerRole-level CSV asking EHR vendors and payers for clinician-to- organization affiliations with a 10-year lookback, included when a clinician billed more than 10 patients under the organizational NPI. Timeliness is explicitly prioritized over completeness.

New pre-registration: H45, the endpoint coverage gap by state

The CEHRT scrape makes a question measurable that we could not measure before. The NDH's weakest layer is endpoints: 98.7% of its organizations carried zero Endpoint references in the April release (H5), and the endpoint records that exist carry an address and payload type while every optional metadata field measures 0% (H44). Meanwhile, thousands of provider organizations already publish a FHIR endpoint through their EHR vendor's HTI-1 bundle, with an NPI and a state attached.

H45 joins the two: for each state, how many organizations with a publicly-published EHR FHIR endpoint either have no matching NDH Organization, or match one that carries zero endpoints? The null hypothesis says the NDH already covers at least 90% of them. The baseline says it will not be close. Both sides of the join are public data, so every per-state count will be independently checkable, and the results will feed the per-state audit pages at /states.

Pre-registration, denominator, and join rules: https://ainpi.dev/findings/cehrt-endpoint-coverage-gap

Also shipped since the last update

  • Per-NPI report cards at https://ainpi.dev/npi: one page per record in the 10,000-record high-risk cohort export, with every signal decoded and primary-source verify links (LEIE, SAM, NPPES) on each page.
  • An MCP server at https://ainpi.dev/api/mcp: five tools (findings, state audits, NPI cohort checks) over the same public contract, so AI agents can query the audit directly. Docs at https://ainpi.dev/developer.
  • The NDH release archive as parquet: both ingested releases (April and May), flattened to the same schema as our BigQuery tables, with DuckDB examples that reproduce published findings exactly. directory.cms.gov serves only the latest release; the archive is how you diff them.
  • A data snack: five software products serve half of the directory's live FHIR hosts, and 36.9% of conformant hosts still declare a superseded FHIR version. https://ainpi.dev/articles/fhir-server-census

Method notes

Repo activity is read from the public GitHub repos (ftrotter-gov, DSACMS); counts of the CEHRT cache are file counts at the 2026-07-10 snapshot, not dedup-adjusted. H45 computes nothing until the join rules in the pre-registration run against a pinned snapshot; the match rate (NPI join vs name-plus-state fallback) will be published alongside the gap counts so the join quality is itself visible.

— Eugene Vestel, FHIR IQ