The integrity disclosure.
Every numeric claim Mohala publishes is reproducible from public filings. This page documents the sources, the joins, the AI policy, and — most importantly — what we don't know yet.
356 tables in HHIP
The Hawaiʻi Health Intelligence Platform — a single audited spine all ten modules read from.
72+ verified sources
Every source registered with cadence, last-success timestamp, and consecutive-failure tracking in data_freshness_watchdog.
Append-only audit chain
audit_evidence_chains records the claim, methodology, and source filing for every published number.
Eight agency families, 72+ feeds, every refresh tracked.
| Agency | Feeds | Cadence |
|---|---|---|
| CMS | Provider Enrollment · Hospital MRF (TiC) · Geographic Variation · RHTP awards | Monthly–Quarterly |
| BLS | OES wages · QCEW employment | Quarterly |
| BRFSS / ACS | Health indicators · Demographic strata | Annual |
| DCCA (HI) | License verification · Enforcement actions | Weekly |
| NPPES | NPI registry & taxonomy | Weekly |
| APCD (HI) | Claims summary · Allowed-amount percentiles · Anomaly detection | Quarterly |
| HMSA / Kaiser / UHA / AlohaCare | Provider directories · Network reachability | Quarterly |
| FEMA / NOAA / USGS | Hazard layers feeding cascade scenarios | Event-driven |
Full per-feed status, last-success timestamps, and failure history live in the operator-facing data_freshness_watchdog table — surfaced to admins under /app/admin-emails.
How disparate filings become one row of evidence.
Specialty normalization
NUCC taxonomy → Mohala 41-specialty grouper. Maps Internal Medicine subspecialties consistently across CMS, NPPES, and insurer directories.
Island/county join
All point data resolved to county FIPS (Honolulu, Hawaiʻi, Maui, Kauai, Kalawao). HPSA scores joined at county-discipline grain.
Entity master crosswalk
EIN + legal-name fuzzy match across CMS, IRS 990, DCCA, and APCD. Parent-child links validated against bond filings before write.
Confidence scoring
Every audit chain carries a 0–1 confidence (sources, recency, cross-source agreement). Exports surface this; UI badges anything <0.7.
AI drafts. Humans publish.
- ·AI is used to draft and summarize — never to invent numeric claims.
- ·Every numeric claim must trace to a source filing in audit_evidence_chains. AI cannot mint new evidence.
- ·Sensitive outputs (testimony, regulator letters) route through ai_output_approvals before publication.
- ·We disclose model and prompt class in PDF footers when AI assistance materially shaped wording.
What we don't know yet.
- Self-pay / cash-pay rates are not in APCD. M10 benchmarks reflect commercial + Medicaid only.
- Federally-run facilities (VA, Tripler) are out of DCCA license scope and partially out of NPPES taxonomy depth.
- NHSC scholarship fulfillment data updates twice a year — between cycles, M6 attrition signals lag actual departures.
- TiC MRF coverage is uneven across small payers; gap is documented per source in data_freshness_watchdog.
Every published claim opens to its evidence chain in two clicks.
Click any numeric claim in the op-ed library to see the underlying chain — sources, methodology, confidence, and the row in HHIP it was derived from.
Browse the op-ed library