AINPI·ainpi.dev
NDH·2026-06-02
Release update · 2026-06-02

Landscape becomes the front door.REAL Health audit framework published.

Landscape cells
548
Audit dimensions
6
Compliance window opens
PY 2028

A hierarchical treemap replaces the choropleth as the AINPI homepage — 548 cells, six dimensions, one cell per state × specialty. Paired with a policy brief that maps every § 6220 obligation to the AINPI signal that measures it.

Read the full report →
methodology v0.7.1-draftNDH 2026-06-02

2026-06-02 update — landscape becomes the front door, REAL Health audit framework published

Two coordinated releases tonight. They are intentionally a pair.

1. / is now the provider data landscape

For the past year the AINPI homepage was a US choropleth. It answered one specific question well: "how many federally-excluded NPIs are still listed in the federal directory, per state?" That question still matters and the map is still live at /map.

But the choropleth was a single-finding view, and AINPI now produces dozens of cross-source signals across millions of records. The homepage needed to match the actual scope of the work.

The new landing page is a provider data landscape — a hierarchical treemap inspired by Andrej Karpathy's US job-market visualizer. Every cell is one (state, specialty) tuple. Area scales with the count of active practitioners. Color is the audit metric in the layer you select.

The five-second pitch: it shows you the federal provider directory the way labor economists show you the workforce — by structure, by mass, and by quality, all in one frame. Switch the layer to flip the color; spatial layout never changes, so you learn the geography once.

Six layers, each measuring a different dimension of accuracy:

  • Completeness — share of records with every § 6220-required field populated
  • Cross-source agreement — share findable in NPPES AND a payer FHIR directory AND PECOS
  • Currency — median days since meta.lastUpdated (90-day cadence in § 6220 fails above 90)
  • Endpoint reachability — share with a live FHIR endpoint on the managing org
  • Federal integrity — share NOT flagged by LEIE / SAM / NPPES-deactivation
  • Specialty validity — share whose NUCC taxonomy resolves cleanly against the CMS Medicare crosswalk

Tile-by toggle on top: group cells by Specialty → state (default) or State → specialty. Same data, different organizing dimension. Click any cell to open a side panel with all six scores side-by-side vs the national baseline, plus a sample of NPIs with primary-source verify links (NPPES Registry, OIG LEIE, SAM.gov) — no API key, no login.

Today the cell-level numbers are a deterministic synthetic seed, clearly marked with a yellow on-page callout. The aggregation script (analysis/landscape.py) is committed and runs against BigQuery with the project's 100 GB per-query cap; the next weekly-refresh cron replaces the seed with measured values. We are shipping the UI first because the visualization design is what makes the audit substrate legible — to regulators, to plans building toward 2028 compliance, to anyone who is about to comment on the CMS scoring-methodology RFC.

2. The REAL Health Providers Act audit framework

HR 7148 § 6220 — the Requiring Enhanced & Accurate Lists of Health Providers Act — was signed into law on 2026-02-03. Medicare Advantage plans must verify every provider record every 90 days, remove departed providers within 5 business days, and submit an annual accuracy analysis to HHS. Starting with plan year 2029, CMS will publish each plan's accuracy score on cms.gov in machine-readable format.

The hard part is not the cadence. It is the measurement methodology. CMS has not yet defined how the 2029 published score is computed. Three paradigms compete:

  1. Field-level on plan-owned data. Easy. Produces scores in the high-90s. Cannot detect a plan's own data being wrong because it is the plan's own definition of correct.
  2. Phone-audit secret shopper. Catches real ghost networks. Senate Finance Committee found 80%+ ghosts in mental-health directories this way. Creates administrative burden, scales poorly, lags reality.
  3. Cross-source intersection. Provider-owned reality (NPPES, PECOS, provider-attested) joined to payer-owned reality (active contract, effective dates, claims observation). Scored at record level against independent sources of truth.

Cross-source decomposed scoring is the only approach a plan cannot grade itself on. It is what AINPI implements. It is what the new landscape view shows. And it is what we believe the 2029 published score should be built from.

The new policy brief at /real-health-providers walks through every § 6220 obligation, maps each one to the existing AINPI signal that measures it, and ships copy-paste citation language for any submitter to the 2028 CMS scoring-methodology RFC. The 18 months between now and the 2028 plan-year compliance window are when the framework gets set. There is a slot for a public, reproducible, citation-grade third-party audit substrate. The pair of releases tonight is the version of AINPI that claims that slot.

What's next

  • Real BQ run on the landscape. Weekly-refresh cron replaces the seed.
  • Per-MA-plan scoreboard. The H26 4-payer probe generalized — one row per (plan, state, specialty) with the six dimensions side-by-side.
  • MCP server. lookup_npi, cross_source_check, get_real_health_score(plan_id, state) — making AINPI the substrate AI agents reach for when answering provider-directory questions.
  • Per-NPI history view. Cross-references each practitioner across NDH releases. The audit instrument for the 5-business-day removal rule.
  • Diff-since-last-release feed. Quarterly NDH releases become the natural compliance-cadence instrument.

Hard constraints we keep

  • No "produced for" or "guided by" any state agency or CMS.
  • No plan or vendor named in any adjective; numbers do the talking.
  • No cost-of-error claims without primary-source verification per record (the 2026-05-22 H40 SAM-NPI false-positive incident remains the cautionary tale).

— Eugene Vestel, FHIR IQ