Datasets I keep coming back to

Facility-level compliance, inspection, and violation history across the Clean Water Act, Clean Air Act, RCRA hazardous waste, and SDWIS drinking water. Every facility EPA tracks is in here with its inspections, violations, formal enforcement actions, penalties, and quarterly significant-violator flags.

powers /the-three-year-list — facilities flagged Significant Violator in 28+ consecutive quarters under CWA with no federal enforcement action.

Self-reported chemical releases from ~21,000 industrial facilities since 1987. Per-facility, per-chemical pounds released to air, water, land, off-site transfer. Coverage is the chemicals on the TRI list at the reporting thresholds.

underbuilt for journalism on local chemical exposure and facility risk maps. The data is structurally clean; the story-shape is the question.

Every recorded tornado, hailstorm, flood, severe-weather event in the US since 1950. Lat/lon (when known), county, casualties, damage estimate. Tornadoes have F/EF rating, path length, path width.

powers tornadolookup.com — the county-level historical tornado index. Spatial-chain reconstruction is required for county-row splits.

Annual workplace-injury filings from employers required to report by OSHA. ~400,000 establishments in the most recent filing year. DART (days-away/restricted/transferred), DAFW (days-away-from-work), TRIR (total recordable injury rate) per establishment, plus the underlying counts and total hours worked. NAICS-tagged.

powers OSHA Lookup — searchable per-establishment safety profile with comparison to industry NAICS-6 median.

A near-census of US employment and wages from the unemployment-insurance system. Industry (NAICS, six-digit where disclosure allows) × area (national, state, county, MSA) × ownership, with employment counts, total wages, number of establishments, average weekly wage. ~75 MB zip per year for the annual single-file.

powers SMB Density — small-business firm density by industry and state.

Median wages, employment counts, and 10/25/50/75/90 percentile wages for ~830 SOC occupations across ~400 MSAs and statewide. Sister-cut to QCEW (industry side) at the occupation side.

could power wages-by-occupation-by-metro — the answer to "what does a plumber actually earn in Phoenix vs Memphis." Queued in the SMB-Density mold.

Individuals and entities barred from receiving federal healthcare-program payments — Medicare, Medicaid, and any program HHS funds. ~70,000 entries with names, exclusion reason, exclusion date, often DOB, often NPI when present. The denominator for any federal-healthcare-fraud anti-join.

powers the LEIE × State-Medicaid investigations track. The 70.9k no-NPI individual subset is 98.8% DOB-populated, which is what makes the cross-table join useful.

Payments from drug and device manufacturers to physicians and teaching hospitals. Mandated by the Sunshine Act since 2013. Per-payment records: drug or device, manufacturer, recipient physician (with NPI), amount, nature of payment, date. Hundreds of millions of dollars per year, individually itemized.

underused for conflict-of-interest reporting at the physician-specialty level. A doctor recommending a specific brand of stent often has structured data in here saying who paid them, when, and for what.

Every healthcare provider with a National Provider Identifier. Individual NPIs (type 1) for physicians, dentists, nurses; organizational NPIs (type 2) for clinics, hospitals, pharmacies. Names, addresses, taxonomy codes (specialty), license info when present.

sister surface to LEIE: when you need to confirm an exclusion record refers to a specific real provider, NPPES is the join.

Safety profiles for every interstate motor carrier with an active DOT number. ~600,000 carriers. BASIC scores (unsafe driving, hours-of-service, vehicle maintenance, crash indicator, drug/alcohol, controlled substances), inspection counts, out-of-service rates, recorded crashes.

powers CarrierLookup — state-ranked carrier safety records. The state-rank filter requires has_sms AND insp_total ≥ 5 to keep one-inspection sole-proprietors from dominating; sole-prop entries stay searchable.

Every FDIC-insured bank branch in the US with deposit total as of June 30 of each year. ~76,000 branches across ~4,500 institutions. Branch-level addresses, deposit balances, institution lineage. Goes back to 1994.

could power bank-desert maps, fintech competitive intel, M&A scouting. Queued in the SMB-Density mold.

Every SBA-backed loan in the 7(a) and 504 programs, going back decades. Per-loan records: lender, borrower, NAICS, amount, term, approval date, status (paid in full, chargeoff, current). ~2–3 million rows total.

could power "what got funded in my industry this year" by NAICS × state. Queued in the SMB-Density mold; the row volume forces an aggregation step before any browser-side tool.

Flood-zone polygons for everywhere FEMA has mapped — Special Flood Hazard Areas, base flood elevations, floodways, levees, coastal zones. Effective version + preliminary/historical layers. The authoritative source for "is this address in a 100-year floodplain."

powers FloodZoneMap — address lookup with adjacent FEMA, USGS, USDA, NOAA, and EPA layers fused at the point.

Every recorded earthquake worldwide. M0+ within continuously-instrumented regions (most of the US), M2.5+ globally. Lat/lon, depth, magnitude, time, focal mechanism where computed. Historical catalog back well before instrumentation for major events.

could power earthquake-near-me + per-county seismic risk pages. On the top-three list of gov-data niches from the 2026-04 portfolio research.

Every recorded US mineral occurrence — producing mines, past-producing mines, prospects, occurrences. Commodities, operators, geology, deposit type. Coverage thins outside the US (sister datasets cover specific regions).

could power mines-near-me with operator-page tree. On the top-three gov-data niches list; the operator-pages-as-third-dimension shape is the key.

Index records (not certificates) of California births, deaths, and marriages. Names, dates, parents (for births), spouse (for marriages), county. Released under the Public Records Act with specific privacy carve-outs.

powers CBI, CDR, and the CA marriage records site. Subject to CCPA opt-out at the individual level; the CBI removal track is part of the operating discipline.

Every US patent ever granted, with full text, claims, citations, classification (IPC + CPC), inventors, assignees. Weekly bulk file from issue date, with retrospective bulk for the entire grant history. PatentsView normalizes it for analysis.

a small-curation cousin powers Patent of the Day. The full corpus is heavy but the slice for any specific thesis (deceptive design patents, expired-coverage workarounds, sleeper assignments) is workable.