Skip to content
American Factbook
Menu

Data

Data

Every dataset this site is built from, with the agency that publishes it, the terms it is published under, how often it changes, the date this copy was taken, and what the copy contains. Where we hold a cleaned table, it is downloadable here.

Each of these is a federal government work and is in the public domain under 17 U.S.C. § 105. The cleaning, normalization, and column naming are ours; take them and use them. We ask only that you cite the original agency rather than us, because the agency is the record and we are a copy of it.

Data retrieved

License

Each dataset below keeps its publisher’s terms, stated in its record. The American Factbook’s own work built from them, its text, charts and derived tables, is licensed CC BY 4.0: reuse it with credit to the American Factbook.

Register

The same register, for a harvester or a program, is a DCAT-US catalog: one entry per dataset, with its downloads, its source and its terms.

Every dataset held here, with its publisher, size, update cadence and the date this copy was fetched.
DatasetRowsDownload
Bureau of Justice Statistics — Capital Punishment2,372CSV, 12 KB
Bureau of Justice Statistics — Jail Inmates224CSV, 2 KB
Bureau of Justice Statistics — Prisoners12,457CSV, 65 KB
Bureau of Justice Statistics — Probation and Parole in the United States14,938CSV, 91 KB
Nationwide Encounters52,986CSV, 293 KB
Census Bureau county population estimates18,864CSV, 50 KB
Census Population Estimates — state characteristics43,992CSV, 259 KB
Hospital General Information5,4194 tables, 2 KB
CMS nursing homes (Provider Information)14,6905 tables, 16 KB
CMS Medicare Advantage and Part D plan landscape130,3074 tables, 38 KB
State Employee Payroll Data Calendar Year 2015 through Present29,880,5256 tables, 859 KB
Connecticut sports wagering — every licensee's wagers, revenue and payment to the state, by month272CSV, 13 KB
Defense Casualty Analysis System — casualty summaries by operation9,590CSV, 32 KB
Equity in Athletics Disclosure Act data — every college athletics program, by institution34,8432 tables, 389 KB
Equity in Athletics Disclosure Act data — every varsity team, by college and sport291,2362 tables, 380 KB
EOIR immigration court outcomes by judge, court and fiscal year139,0474 tables, 70 KB
EOIR immigration judges and courts12,370CSV, 87 KB
FBI Crime Data Explorer — agencies and the population each serves19,636CSV, 410 KB
FBI Crime Data Explorer — crime reported by each agency195,276CSV, 1.4 MB
FBI Crime Data Explorer — Arrests45,500CSV, 225 KB
FBI Crime Data Explorer API832,124CSV, 4.0 MB
FBI Crime Data Explorer — Hate Crime106,964CSV, 379 KB
FBI Crime Data Explorer — Police Employment and LEOKA2,362CSV, 12 KB
FBI Crime Data Explorer — Supplementary Homicide Reports288,288CSV, 1.1 MB
FDIC failures — every bank failure and assistance transaction since 19344,1192 tables, 2 KB
FDIC BankFind — institutions (every charter the FDIC has insured, active and inactive)27,8352 tables, 6 KB
FDIC BankFind — locations (every current office of an FDIC-insured institution)78,0772 tables, 3 KB
FDIC Summary of Deposits — every office's deposits at June 30, 1994 to the newest survey2,898,8362 tables, 68 KB
FDIC summary — insured institutions, offices, assets and deposits by state and year since 19347,7612 tables, 111 KB
FEMA's FIFA World Cup Grant Program — every award under assistance listing 97.1609CSV, 1 KB
Farm program payment files77,404,4269 tables, 195 KB
OPM Federal Workforce Data — Federal Employment Raw Data (EHRI Status)432,907CSV, 905 KB
U.S. House of Representatives — Statement of Disbursements (payroll)698,0303 tables, 4 KB
IPEDS — every postsecondary institution, with NCES's derived headline measures5,985CSV, 4 KB
Form 990 series e-file data760,59214 tables, 351 KB
Exempt Organizations Business Master File1,975,9756 tables, 6.0 MB
State of Minnesota — State Employee Payroll Data2,215,1384 tables, 59 KB
NCES Common Core of Data — public school districts, newest full school year19,6302 tables, 348 KB
NCES School District Finance Survey (F-33) — newest fiscal year19,570CSV, 628 KB
NCES School District Finance Survey (F-33) — every fiscal year in the current format367,012CSV, 45 KB
NCES Common Core of Data — public schools, newest full school year102,1782 tables, 5 KB
NCES Common Core of Data — public school membership by state, 2014–15 onwards11,080CSV, 8 KB
NCUA credit-union sites — Credit Union Branch Information (the newest quarter)22,837CSV, 1 KB
NCUA 5300 Call Report — quarterly data (every credit union, every quarter from March 1994)860,4873 tables, 14 KB
State of New Jersey — YourMoney Agency Payroll1,380,0365 tables, 42 KB
State of Oklahoma Payroll29,579,8573 tables, 547 KB
Open Payments — payments grouped by physician7,458,1739 tables, 47 KB
Open Payments — payments grouped by covered recipient and nature of payment7,300,1162 tables, 2 KB
OPM FedScope Employment Summary Data14,511CSV, 183 KB
Medicare Part D Prescribers — by provider1,416,8837 tables, 27 KB
SBA Paycheck Protection Program (PPP) FOIA — loan-level data11,468,2108 tables, 259 KB
Treasury Fiscal Data97,779CSV, 1.0 MB
USAspending API v28,351CSV, 100 KB
USAspending Award Data Archive — contracts2,404,8315 tables, 140 KB
USAspending API v2 — dimension aggregations210,485CSV, 666 KB

Row counts are of the copy held here, which for some sources is a subset of what the agency publishes. Where that is so it is stated below and the whole file is named.

Bureau of Justice Statistics — Capital Punishment

Publisher
U.S. Department of Justice, Bureau of Justice Statistics
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual
This copy fetched
Rows held
2,372
Columns
8 columnsedition, table_id, scope, jurisdiction, year, measure, category, value
Download
Capital punishment tables, as published — gzipped CSV, 12 KB

What is committed here is a summary. The full table has 2,372 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 7d757d69b3867b372e6b9fcd5050bb12fcd931e0231afe1cca53d9274be4c10a, so every figure here can be traced to the bytes it came from.

What it cannot show

— 7 limitations
  • A prisoner under sentence of death on 31 December is a person whose sentence stood on that date. It is not a count of people who will be executed: most sentences end in a court overturning them, a resentencing, or the prisoner dying of other causes, and the tables count those removals separately from executions.
  • Executions are counted in the jurisdiction that carried them out, and the cumulative figures run from 1977 — the year after Gregg v. Georgia allowed the penalty to resume. The national series runs from 1930 and the two are never added together.
  • A jurisdiction with no death penalty statute can still hold prisoners under a sentence imposed before it was repealed. A zero in one column and a positive number in another for the same state is that, not an error.
  • The regional rows in the by-jurisdiction tables are subtotals of the states beneath them. Summing a region with its states double-counts, so the scope of every row is recorded rather than inferred from its position.
  • Race and ethnicity in table 2 are on the definition used from the 2023 edition, where white and black exclude prisoners of Hispanic origin. The long series in appendix table 3 uses the older definition, where they do not: 1,243 white prisoners in 2023 against table 2's 935. Both are published, under different category names, and they are not two points on one series.
  • The receipts and removals in appendix table 2 do not reconstruct the published population series, in 46 of 51 years, by up to 19 prisoners. The population is revised as late reports arrive and the flow series is not.
  • These are state prison systems and the federal system. Military executions and any carried out by a jurisdiction outside that frame are not here.

Bureau of Justice Statistics — Jail Inmates

Publisher
U.S. Department of Justice, Bureau of Justice Statistics
License
Public domain. The Department of Justice's reuse policy states: "Information generated by the Department of Justice is in the public domain and may be reproduced, published or otherwise used without the Department's permission." It asks that the Department not be represented as endorsing any product or activity, and that it be acknowledged as the source. https://bjs.ojp.gov/reuse-policy
Cadence
Annual
This copy fetched
Rows held
224
Columns
7 columnsedition, scope, jurisdiction, year, measure, value, coverage
Download
Jail populations, as published — gzipped CSV, 2 KB

What it cannot show

— 6 limitations
  • A jail is not a prison. Jails are locally operated and hold people awaiting trial and those serving short sentences; prisons are run by a state or the federal government. The two are separate collections and are never added, because a person can appear in both in one year.
  • Six states run combined jail and prison systems and so have no locally operated jails to count: Connecticut, Delaware, Hawaii, Rhode Island, Vermont, and West Virginia, which was in the 2019 tables and is absent from 2024. They are missing from the state table rather than reporting zero.
  • Alaska appears in the table and its row is partial. It runs a combined system and also fourteen locally operated jails, and the row counts only those fourteen — 31 people, a rate of 4 per 100,000. It is flagged and excluded from rankings, because presenting it beside a complete state would be wrong by two orders of magnitude.
  • The state figures come from a Census of Jails year, which is a complete enumeration. In other years BJS surveys a sample and publishes national estimates with standard errors but nothing by state, so a state series is a handful of points rather than an annual trend.
  • The rate is BJS's own, per 100,000 U.S. residents of all ages, computed against its own denominators. Nothing here divides a count by a population of ours.
  • The Census of Jails 2005-2019 state tables are not ingested. They use a third header shape, with years as column labels beneath measure bands and percent-change columns between them, and one guessed column there would publish a rate as a count.

Bureau of Justice Statistics — Prisoners

Publisher
U.S. Department of Justice, Bureau of Justice Statistics
License
"Information generated by the Department of Justice is in the public domain and may be reproduced, published or otherwise used without the Department's permission." (https://bjs.ojp.gov/reuse-policy)
Cadence
Annual, published roughly two years after the reference year
This copy fetched
Rows held
12,457
Columns
6 columnsedition, table_id, jurisdiction, year, measure, value
Download
Prison populations, as published — gzipped CSV, 65 KB

What it cannot show

— 6 limitations
  • Jurisdiction and custody are different counts and are never interchangeable. Jurisdiction is the legal authority of a correctional agency over a prisoner wherever that prisoner is held; custody is where the prisoner physically is. A state's jurisdiction count includes people it houses in private prisons and local jails.
  • The per-jurisdiction tables in one edition carry two reference years, not the full series. A state-level trend is assembled by ingesting several editions, so a state series is only as long as the number of editions held.
  • BJS revises earlier years in later editions. Where two editions disagree about the same jurisdiction and year, the later edition is the published figure and the earlier one is kept in the snapshot rather than overwritten.
  • Imprisonment rates are BJS's own, computed against its own resident-population denominators. This pipeline stores the rate BJS published rather than dividing counts by a population of our own.
  • Counts are as of December 31 of the reference year except for federal prisoners in some tables, which BJS reports as of September 30. The tables that differ say so.
  • Some jurisdictions do not report every item. BJS marks these, and a value it did not publish is stored as null rather than as zero.

Bureau of Justice Statistics — Probation and Parole in the United States

Publisher
U.S. Department of Justice, Bureau of Justice Statistics
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual
This copy fetched
Rows held
14,938
Columns
11 columnsedition, year, table_id, supervision, subset, scope, jurisdiction, measure, category_group, category, value
Download
Probation and parole tables, as published — gzipped CSV, 91 KB

What is committed here is a summary. The full table has 14,938 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is c2224a317df144d650f7be7f44cf7ca228f0b185fbed7ff5cad69f7c88fdf6d4, so every figure here can be traced to the bytes it came from.

What it cannot show

— 6 limitations
  • The surveys count people under supervision, not people. A person on probation in two states is two records, and a person who leaves and returns within a year appears in both entries and exits. Nothing here is a count of individuals.
  • Entries and exits are published twice: as reported by the agencies that answered, and imputed for those that did not. The two are loaded separately and are never added. Start plus imputed entries minus imputed exits reconstructs the December population to within about 2%, which is the size of the imputation rather than an error.
  • The State row is the subtotal of the fifty states and the District of Columbia, and the “U.S. total” row adds the federal system to it. Summing State with the states beneath it double-counts, so the scope of every row is recorded rather than inferred.
  • Maine abolished parole in 1976 and is absent from the parole tables after 2021. An absent jurisdiction is absent, not zero, and .., /, ~ and a suppressed <10 are all loaded as null for the same reason.
  • The editions revise each other. The 2023 edition republishes parole for 2019, 2020 and 2021, and its 2021 figures differ from the 2021 edition's — California's parole population was revised from 119,595 to 110,752, which moves the national total by 8,843. Both are loaded, under the edition that published them, and a trend that crosses a revision says so rather than drawing a line through it.
  • Race and ethnicity are as the supervising agency recorded them, and a quarter of the probation population is “Unknown/not reported” in some years. A share computed over the known cases is a share of what was recorded, not of the population.

Nationwide Encounters

Publisher
U.S. Customs and Border Protection
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Monthly
This copy fetched
Rows held
52,986
Columns
12 columnsfiscal_year, fiscal_year_partial, month_abbr, month_number, component, land_border_region, area_of_responsibility, demographic, citizenship, title_of_authority, encounter_type, encounter_count
Pages built on it
Immigration and the border
Download
Encounters by month, sector and demographic — gzipped CSV, 293 KB

What it cannot show

— 5 limitations
  • An encounter is an event, not a person. The same person may be encountered more than once in a period, so encounter counts are not a count of individuals.
  • Encounter counts describe enforcement activity. They do not measure migration, the number of people who crossed, or the size of any population.
  • Title 8 and Title 42 actions are different legal authorities with different consequences. Adding them together without saying so misrepresents both.
  • CBP republishes the file each month and revises earlier months in it. A figure is only as current as the snapshot it was drawn from.
  • "Area of Responsibility" is the CBP sector or field office that recorded the encounter, not the place a person entered or intended to go.

Census Bureau county population estimates

Publisher
U.S. Census Bureau
License
U.S. Government work, and the publisher's own words rather than recalled ones. The Census Bureau's Open Government page says: "The U.S. Census Bureau publishes its data as open data, meaning it is freely available for use and re-use by the public" — https://www.census.gov/about/policies/open-gov.html, read 23 September 2026. Its citation policy asks to be credited: "Proper citation ensures that Census Bureau statistical products and research can be discovered, reused, replicated for verification, and credited for recognition" — https://www.census.gov/about/policies/citation.html, same date. That is honoured rather than treated as optional: every page carrying a county figure from this file names the U.S. Census Bureau.
Cadence
Annual
This copy fetched
Rows held
18,864
Columns
8 columnsfips, state_fips, state_name, state_abbr, county_name, county_key, population_year, population
Download
Counties, with their FIPS codes — gzipped CSV, 50 KB

What is committed here is a summary. The full table has 18,864 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 6cdad8e9aa02beee42215da01331172b8baa27da4c8be545abf5537be5cd5caa, so every figure here can be traced to the bytes it came from.

What it cannot show

— 5 limitations
  • The Bureau publishes no license identifier and no page of terms. What it publishes is the open-data sentence quoted above and a citation policy. The public-domain position for a U.S. Government work is 17 U.S.C. 105, which is a statement about federal works rather than a quotation from this publisher; recording it as the publisher's words is a mistake this site made once and has corrected.
  • A county is the Census Bureau's county, today. Connecticut's counties were replaced by nine planning regions in the 2022 vintage and the current file has the regions; publishers that still use Connecticut's counties — USDA's Farm Service Agency among them — will not join against it. That is a fact about two publishers rather than a defect in either, and the page that joins says which rows did not.
  • A county equivalent is not always a county. Louisiana has parishes, Alaska has boroughs and census areas, Virginia's independent cities are separate entities from the counties that surround them, and Puerto Rico has municipios. All are in this file, under their own names.
  • The population is an estimate between decennial censuses, revised every year. A figure from the 2025 vintage and one from the 2023 vintage for the same county in the same year are frequently different, and neither is a count of people.
  • The state rows — “COUNTY = '000'” — are the state's own total and are not counties. They are dropped here rather than published, because a table that mixes a state with its counties double-counts every time somebody sums it.

Census Population Estimates — state characteristics

Publisher
U.S. Census Bureau
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual, released the June after the reference year
This copy fetched
Rows held
43,992
Columns
6 columnsstate_abbr, year, dimension, category, population, vintage
Download
Population by state, county and year — gzipped CSV, 259 KB

What it cannot show

— 5 limitations
  • The FBI's race categories and the Census Bureau's are not the same categories. The FBI records one race per arrestee, assigned by the arresting agency; the Census records race as the person reports it and allows more than one. A rate built from the two is an arrest count over a population that was counted a different way, and every page that shows one says so.
  • The FBI has no ethnicity field, so no Hispanic or Latino rate can be computed. The denominators here are therefore race alone across all origins — never race alone non-Hispanic, which would divide a numerator that includes Hispanic arrestees by a population that excludes them and inflate the rate.
  • There is no Census equivalent of the FBI's "Multiple" race category. Census "Two or more races" is self-identified and the FBI's is an officer's classification of one person. No denominator is published for it, so no rate is.
  • Population estimates are estimates, revised each vintage. A rate is only as current as the vintage its denominator came from, which is recorded with every row.
  • These are state figures. County population crossed with race is published only in five-year age bands, which do not align with the FBI's bands below age 25, so no county rate is derived for those bands.

Hospital General Information

Publisher
Centers for Medicare & Medicaid Services
License
U.S. Government work, in the publisher's own words rather than recalled ones. The Provider Data Catalog's own machine-readable catalog, https://data.cms.gov/provider-data/data.json (DCAT-US 3.0, read 27 September 2026), lists this dataset and 236 others and carries no "license" field on any of them. CMS's other catalog, https://data.cms.gov/data.json (Project Open Data v1.1), lists 159 datasets, every one "license": "https://www.usa.gov/government-works", and none of them this file. So, as for the Medicare plan landscape, the statement that covers it is CMS's website policy: "CMS.gov is the official website of the Centers for Medicare & Medicaid Services (CMS), an operating division of the U.S. Department of Health & Human Services (HHS). It is a public domain web site." — https://www.cms.gov/about-cms/web-policies-important-links/about-website/link-to-us. The condition the government-works page carries is the site's standing rule: no CMS seal, no logo, and no page calling itself official Medicare information; CMS is named as the publisher and the dataset linked on every page. Decided by Kian on September 27, 2026: "same footing as the Medicare plan landscape".
Cadence
Quarterly
This copy fetched
Rows held
5,419
Columns
39 columnsccn, facility_name, address, city, state, zip_code, county, telephone, hospital_type, ownership, emergency_services, birthing_friendly, overall_rating, overall_rating_footnote, mortality_group_measures, mortality_measures, mortality_better, mortality_no_different, mortality_worse, mortality_footnote, safety_group_measures, safety_measures, safety_better, safety_no_different, safety_worse, safety_footnote, readmission_group_measures, readmission_measures, readmission_better, readmission_no_different, readmission_worse, readmission_footnote, patient_experience_group_measures, patient_experience_measures, patient_experience_footnote, timeliness_group_measures, timeliness_measures, timeliness_footnote, released

What is committed here is a summary. The full table has 5,419 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is c45a59684520d3a44a2c338304c95e7c2b75dffbcbcc4375980e348c44934740, so every figure here can be traced to the bytes it came from.

Every committed table for Hospital General Information, with its size and a download.
TableRowsSizeDownload
Hospitals by ownership121 KBCSV
Hospitals by state561 KBCSV
Hospitals by type80 KBCSV
Hospitals nationally10 KBCSV

What it cannot show

— 4 limitations
  • A rating is CMS's, and many hospitals have none. 2,245 of 5,419 hospitals had no overall rating in the July 2026 file, each with CMS's footnote giving the reason (too few measures reported, a hospital type CMS does not rate, and others). A missing rating is shown as CMS's reason, never as a zero or an average.
  • "Better", "no different" and "worse" are CMS's comparisons with the national rate, made measure by measure within each group. They are counts of measures, not a score, and a hospital reporting few measures has few comparisons.
  • Veterans Affairs, Department of Defense and some other hospitals report few or no measures. 850 hospitals carry "Not Available" for every measure group; their pages say so rather than printing zeros.
  • This file carries no county FIPS code, only the county's name, so every join to a county page is a name join within one state. Nothing is linked on a guess.

CMS nursing homes (Provider Information)

Publisher
Centers for Medicare & Medicaid Services
License
U.S. Government work, in the publisher's own words rather than recalled ones. The Provider Data Catalog's own machine-readable catalog, https://data.cms.gov/provider-data/data.json (DCAT-US 3.0, read 27 September 2026), lists this dataset and 236 others and carries no "license" field on any of them. CMS's other catalog, https://data.cms.gov/data.json (Project Open Data v1.1), lists 159 datasets, every one "license": "https://www.usa.gov/government-works", and none of them this file. So, as for the Medicare plan landscape, the statement that covers it is CMS's website policy: "CMS.gov is the official website of the Centers for Medicare & Medicaid Services (CMS), an operating division of the U.S. Department of Health & Human Services (HHS). It is a public domain web site." — https://www.cms.gov/about-cms/web-policies-important-links/about-website/link-to-us. The condition the government-works page carries is the site's standing rule: no CMS seal, no logo, and no page calling itself official Medicare information; CMS is named as the publisher and the dataset linked on every page. Decided by Kian on September 27, 2026: "same footing as the Medicare plan landscape".
Cadence
Monthly
This copy fetched
Rows held
14,690
Columns
51 columnsccn, provider_name, address, city, state, zip_code, county, ownership_type, certified_beds, average_residents_per_day, provider_type, resides_in_hospital, legal_business_name, date_first_approved, chain_name, chain_id, facilities_in_chain, chain_average_overall_rating, continuing_care_retirement_community, special_focus_status, abuse_icon, inspection_over_two_years_ago, changed_ownership_last_12_months, overall_rating, overall_rating_footnote, health_inspection_rating, health_inspection_rating_footnote, quality_measure_rating, quality_measure_rating_footnote, staffing_rating, staffing_rating_footnote, reported_nurse_aide_hours, reported_lpn_hours, reported_rn_hours, reported_total_nurse_hours, weekend_total_nurse_hours, adjusted_rn_hours, adjusted_total_nurse_hours, total_nursing_staff_turnover, registered_nurse_turnover, administrators_left, health_deficiencies_cycle_1, total_weighted_health_survey_score, infection_control_citations, number_of_fines, fines_dollars, payment_denials, total_penalties, latitude, longitude, processing_date

What is committed here is a summary. The full table has 14,690 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is ba34fec3776578b8f73f7e6d6fee69fe77bddae93fe7831f2d02aa14557c2b47, so every figure here can be traced to the bytes it came from.

Every committed table for CMS nursing homes (Provider Information), with its size and a download.
TableRowsSizeDownload
Nursing homes by ownership131 KBCSV
Nursing homes by state532 KBCSV
Nursing home chains61110 KBCSV
The most-fined nursing homes1003 KBCSV
Nursing homes nationally10 KBCSV

What it cannot show

— 5 limitations
  • A rating is CMS's, and some homes have none. 127 of 14,690 homes had no overall rating in August 2026, each with CMS's footnote giving the reason (a new home, too few inspections, a Special Focus Facility). A missing rating is shown as CMS's reason, never as a zero or an average.
  • Staffing is reported by the home. The staffing hours come from the payroll data homes submit to CMS; the adjusted figures are CMS's own case-mix adjustment of them. Neither is an independent count.
  • Fines and penalties cover CMS's look-back period, not the home's history. A home with no fine in this file may have been fined before the period CMS counts.
  • The abuse icon is CMS's flag, with CMS's definition. It marks a home cited for abuse that harmed a resident, or for potential harm, in the period CMS states. It is reproduced as published and is not a finding of this site.
  • This file carries no county FIPS code, only the county's name, so every join to a county page is a name join within one state. Nothing is linked on a guess.

CMS Medicare Advantage and Part D plan landscape

Publisher
Centers for Medicare & Medicaid Services
License
U.S. Government work, in the publisher's own words rather than recalled ones. CMS states on its own website-policy page: "CMS.gov is the official website of the Centers for Medicare & Medicaid Services (CMS), an operating division of the U.S. Department of Health & Human Services (HHS). It is a public domain web site." — https://www.cms.gov/about-cms/web-policies-important-links/about-website/link-to-us, read 26 September 2026. The landscape file is served from that site, from a directory CMS's robots.txt (https://www.cms.gov/robots.txt) does not close to automated retrieval: it closes the site's administrative, account, content-management and Spanish-language sections, and not its file downloads, so nothing forbids fetching the file. This file is not in CMS's machine-readable catalog: https://data.cms.gov/v1-1-data.json answers 200, conforms to the Project Open Data v1.1 schema and lists 159 datasets, every one of them "license": "https://www.usa.gov/government-works", and none of them the landscape (read 26 September 2026). So the catalog's declaration is the publisher's consistent position rather than a statement about this file, and the site-wide public-domain sentence above is the one that covers it. The condition that page's license carries is honored either way and is the site's standing rule: no CMS seal, no logo, and no page calling itself official Medicare information. CMS is named as the publisher and the file is linked on every page built on it.
Cadence
Annual
This copy fetched
Rows held
130,307
Columns
74 columnscontract_year, contract_category_type, us_territory, state_abbr, state_name, county_name, contract_id, plan_id, segment_id, contract_plan_id, contract_plan_segment_id, sanctioned_plan, parent_organization_name, contract_name, organization_marketing_name, organization_type, plan_name, plan_type, snp_indicator, snp_type, snp_institutional_type, snp_institutional_category, dsnp_integration_status, dsnp_aip_identifier, csnp_condition_type, zero_dollar_cost_sharing_dsnp, part_d_coverage_indicator, national_pdp, drug_benefit_category, drug_benefit_type, voluntary_de_minimis, part_d_basic_premium_at_or_below_benchmark, lis_auto_enrollment, offers_drug_tier_with_no_deductible, annual_part_d_deductible, annual_part_d_deductible_note, part_d_basic_premium, part_d_basic_premium_note, part_d_supplemental_premium, part_d_supplemental_premium_note, part_d_total_premium, part_d_total_premium_note, lips_amount, lips_amount_note, part_d_lips_cms_pays, part_d_lips_cms_pays_note, part_d_low_income_beneficiary_premium, part_d_low_income_beneficiary_premium_note, part_d_oop_threshold, part_d_oop_threshold_note, part_c_premium, part_c_premium_note, monthly_consolidated_premium, monthly_consolidated_premium_note, in_network_moop, in_network_moop_note, part_c_summary_star_rating, part_c_summary_star_rating_note, part_d_summary_star_rating, part_d_summary_star_rating_note, overall_star_rating, overall_star_rating_note, ma_region_code, ma_region, pdp_region_code, pdp_region, data_last_updated, prior_contract_year, prior_part_c_summary_star_rating, prior_part_c_summary_star_rating_note, prior_part_d_summary_star_rating, prior_part_d_summary_star_rating_note, prior_overall_star_rating, prior_overall_star_rating_note
Pages built on it
Health

What is committed here is a summary. The full table has 130,307 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 338c3e932158aba5a9aeab6e545e598923c916b802365ac3f5648abd7019eac2, so every figure here can be traced to the bytes it came from.

Every committed table for CMS Medicare Advantage and Part D plan landscape, with its size and a download.
TableRowsSizeDownload
Medicare Advantage plans by county3,21830 KBCSV
Medicare plans by state561 KBCSV
Stand-alone Part D drug plans4767 KBCSV
Medicare plans, national10 KBCSV

What it cannot show

— 8 limitations
  • Not every plan is in this file. CMS excludes employer-sponsored plans (a plan ID of 800 or above), Part B only plans and National PACE plans, and from CY2026 it excludes Medicare-Medicaid Plans, whose demonstration ended in 2025. So a count of plans from this file is a count of the plans CMS lists for a general audience, not of every plan that exists.
  • A stand-alone drug plan is listed once for a state, not for each county. The 549 stand-alone Prescription Drug Plan rows in CY2026 carry the county name "All Counties" and a Part D region, because a PDP is sold across a region rather than by county. A county page that read only its own county's rows would show no drug-only plan at all, so the pages expand a state's PDP rows into each of its counties and label them as sold statewide by region.
  • Premiums can be negative, and that is the record. 23,392 of 138,260 CY2026 rows carry a negative Part D basic premium, written in accounting parentheses — ($19.60). CMS's own readme says the figure "is the net of any Part A/B rebates applied to 'buy down' the drug premium for Medicare Advantage plans", so a negative premium is a rebate exceeding the basic premium and is kept as published, never floored at zero.
  • A star rating is not always a number. In CY2026, 7,696 rows read "Plan Too New To Be Measured", 4,543 "Not Enough Data Available" and 702 "Not Applicable" for the overall rating. Each is kept as CMS's own phrase beside a null rating, never as a zero and never as an average.
  • The figures are as approved, and they change. CMS's readme says "The data within are subject to change as contracts are finalized", and the CY2026 file was revised seven times between its release on 26 September 2025 and 2 September 2026 — sanctions added and removed, and the overall star ratings updated twice. Every page states the contract year and the file's own "Data Last Updated" date.
  • A plan under sanction is marked, and nothing here says what the sanction is. 110 CY2026 rows carry a sanctioned plan flag. The flag is reproduced; the reason is not in this file.
  • Connecticut is published by its former counties. CMS lists Connecticut plans under the eight counties the state used before 2022. The Census Bureau now publishes nine planning regions instead, so those eight have no county page on this site and no population to divide by. The rows are kept under the geography CMS publishes, and the pages say so.
  • This file carries no county FIPS code, only a state abbreviation and a county name, so every join to a county here is a name join within one state. Nothing is linked on a guess.

State Employee Payroll Data Calendar Year 2015 through Present

Publisher
Connecticut Office of the State Comptroller
License
The dataset's own metadata (https://data.ct.gov/api/views/9m78-yc88.json, read 26 September 2026) states "license": {"name": "Public Domain"}, "licenseId": "PUBLIC_DOMAIN", "attribution": "Office of the State Comptroller", with no license link. The portal's Terms of Use (https://data.ct.gov/terms, linked from its footer, read the same day) grant no license of their own and restrict no reuse; they read, in part: "Publishable data made available on the Connecticut Open Data Portal shall be provided as a public service on an “as is” basis. Although the State of Connecticut will strive to ensure that such publishable data are accurate, the State shall make no warranty, representation or guaranty of any type as to the content, accuracy, timeliness, completeness or fitness for any particular purpose or use of any publishable data provided on such portal" and "The State of Connecticut shall reserve the right to discontinue availability of content on the Connecticut Open Data Portal at any time and for any reason."
Cadence
Bi-weekly
This copy fetched
Rows held
29,880,525
Columns
38 columnscalendar_year, payroll_fiscal_year, check_number, check_date, check_option, check_status, published_copy, duplicate_of_published_row, employee_record_id, first_name, middle_initial, last_name, name_suffix, dept_id, agency, department, job_title, employee_class, job_indicator, full_part, union_name, annual_rate, biweekly_comp_rate, salaries_wages, overtime, other_earnings, gross_pay, fringe, fringe_no_retirement, sers_amount, arp_amount, teachers_amount, judges_amount, original_hire_date, termination_date, work_city, work_postal, work_state

What is committed here is a summary. The full table has 29,880,525 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is f07db675b247a6febac91e8136e6614eda76f31612cf1c68ede4cb549fc5e823, so every figure here can be traced to the bytes it came from.

Every committed table for State Employee Payroll Data Calendar Year 2015 through Present, with its size and a download.
TableRowsSizeDownload
Gross pay and employment records by agency in each calendar year1,40649 KBCSV
Printed names paid by agency in each calendar year, and their median pay1,40619 KBCSV
Printed names paid by agency and job title in each calendar year, for the titles with a page50,155628 KBCSV
Printed names paid by job title in each calendar year, titles with ten or more12,589161 KBCSV
Printed names paid in each calendar year, and their median pay170 KBCSV
Gross pay, overtime and employment records in each calendar year171 KBCSV

What it cannot show

— 7 limitations
  • A row is one line of one pay check, not a salary. A person's pay for a year is the sum of their rows for that calendar year. annual_rate is the rate printed on the check — the publisher's own description is "Bi-weekly comp rate times 26.1 for bi-weekly employees, 24 for semi-monthly employees, 12 for monthly employees - If available/may not be current" — and it is not pay received.
  • A reversed check stays in the file. The original carries check_status R and keeps its positive amounts; the reversal is a separate row with check_option R and negative amounts. Summing a person's rows nets them. In 2011, 2012, 2014 and 2015 a few reversals (38, measured 26 September 2026) are split across two payroll fiscal years and print the whole check's gross_pay on each part, so their gross_pay does not net exactly; salaries_wages does.
  • Names are withheld for about one row in eight. The publisher prints "Undisclosed" in every name field of 3,907,022 of 29,878,776 rows (26 September 2026), almost all of them at the University of Connecticut, UConn Health and the Board of Regents. Those rows are kept, with their pay, and the name is carried as printed.
  • employee_record_id is one employment record, not one person. It is the publisher's hashed "EmplId-Empl Rcd": a person with two jobs has two, and nothing the publisher documents joins them. A name is not an identity either; two people can share one.
  • The publisher lists some rows twice. 72,824 rows of 2026, byte for byte, 69,386 of them the ACH payroll of 6 March 2026 — $232,021,311.30 of gross pay counted twice in the portal's own total (measured 26 September 2026). Kian, 26 September 2026: "load as published, mark the duplicate and exclude it from every total, with the note on the state's page". So both copies are in the table, the second marked duplicate_of_published_row, and every total leaves it out: this site's 2026 figures are below the portal's by exactly those rows and dollars, and the state's page says why. The snapshots read on 12 October 2026 had none: every year's published rows equal the rows kept, 2026's 1,356,431 included (extract.yml run 38277056029), so the publisher had removed the repeats; the rule stands for any it prints again, and a page notes a year's repeats only where there are some.
  • gross_pay is the publisher's total and is not always the sum of salaries_wages, overtime and other_earnings: 22,954 of 2025's 1,815,650 rows differ, and the publisher does not say what the difference is. Fringe is the employer's cost, not pay to the employee. The dataset is the State's own payroll — its agencies, the Judicial Branch, the Legislature and the public universities among them — not the payrolls of Connecticut's towns or school districts.
  • The title says 2015, and the file runs from calendar 2010. work_city, work_postal and work_state are where the employee works, not where they live; the publisher publishes no home address.

Connecticut sports wagering — every licensee's wagers, revenue and payment to the state, by month

Publisher
Connecticut Department of Consumer Protection, Gaming Division
License
Public Domain, as the publisher marks it. The portal's own record for both datasets, read 26 September 2026 and on every run, names the license "Public Domain" and the attribution "Department of Consumer Protection"; the portal states the mark on each dataset's own page, https://data.ct.gov/d/xf6g-659c (online) and https://data.ct.gov/d/yb54-t38r (retail), and gives no URL for it. The portal's policies page (https://portal.ct.gov/Policies/State-Privacy-Policy) states no copyright. A run stops if either dataset is no longer marked Public Domain. The Department of Consumer Protection is named as the publisher on every page.
Cadence
Monthly. Licensees pay by the 15th of the following month and the portal is updated soon after: the August 2026 rows were posted on 23 September 2026.
This copy fetched
Rows held
272
Columns
16 columnsmonth, channel, licensee, fiscal_year, wagers, patron_winnings, canceled_wagers, resettlements, win, federal_excise_tax, unadjusted_revenue, promotional_credits_wagered, promotional_deduction, gross_gaming_revenue, payment, notes
Pages built on it
Sports
Download
Every Connecticut sports-wagering licensee, by month and channel — gzipped CSV, 13 KB

What is committed here is a summary. The full table has 272 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 3ab7cb1d5d63298d50c9a0fc68cb15d9b013e9ea9a716f979d5d355cb72e3207, so every figure here can be traced to the bytes it came from.

What it cannot show

— 5 limitations
  • Connecticut's figures only, in Connecticut's definitions. Wagers are the amount wagered; gross gaming revenue is the win less federal excise tax and a promotional deduction the law caps (25% of the win to September 2022, 20% to September 2023, 15% since). Other states define handle, revenue and tax differently, and no figure here is compared with another state's without its label.
  • The payment is not a tax on wagers. It is 13.75% of gross gaming revenue, and nothing in a month when revenue is negative.
  • Months are restated in place. A licensee's amended return replaces the month's figures, as the metadata's notes (8) to (11) record.
  • The figures do not always add up to the win. In six of 272 rows the published win differs from wagers less winnings, canceled wagers and resettlements: five by two dollars, and CT Lottery's retail row for September 2023, the month its XL Center location opened, by $1,215,619. The published win is kept as published.
  • Licensees are the state's. Online wagering runs through three licenses (the tribal operators' and the Connecticut Lottery's), each naming its licensee, not the sportsbook brand. Retail is the Connecticut Lottery's locations and, from September 2023, its XL Center location, whose payments go to the Capital Region Development Authority.

Defense Casualty Analysis System — casualty summaries by operation

Publisher
U.S. Department of Defense, Defense Manpower Data Center
License
U.S. Government work, public domain; https://www.usa.gov/government-works. DoD publishes no terms page for DCAS.
Cadence
Updated continuously; the publisher stamps each table with an extraction date
This copy fetched
Rows held
9,590
Columns
9 columnsconflict, conflict_name, table_id, month, measure, category, service, value, extraction_date
Pages built on it
Defense
Download
Casualties by operation, month and category — gzipped CSV, 32 KB

What is committed here is a summary. The full table has 9,590 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is b15ba9e7914a39ee123781fd3e9eaa3ef67811cfc739b212e6e245fe4a84e55c, so every figure here can be traced to the bytes it came from.

What it cannot show

— 6 limitations
  • These are the operations the Department of Defense names and reports casualties for: Iraqi Freedom, Enduring Freedom, New Dawn, Inherent Resolve, Freedom's Sentinel and Epic Fury. An operation is a named military campaign, not a declaration of war, and a death outside a named operation is not in these figures.
  • Hostile deaths include killed in action, died of wounds, and died while missing, captured or detained. Non-hostile deaths include accidents, illness, homicide and self-inflicted deaths. The categories are the Department's, and the figures say who died and how the Department classified it, not why.
  • Wounded in action counts are reported by month and service only. The publisher does not break them down by category.
  • The monthly and category tables count the same deaths two ways. Every row carries a table_id, and a figure must come from one table or the other, never both added together.
  • One death in Operation Iraqi Freedom has no month in the publisher's monthly table. It is in the operation's total, and it is loaded with no month rather than assigned one.
  • The publisher revises its tables as records are completed. Each row carries the extraction date the publisher stamped on the table it came from.

Equity in Athletics Disclosure Act data — every college athletics program, by institution

Publisher
Office of Postsecondary Education, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works, recorded in our words and not the Department's; decided by Kian O'Connor on 26 September 2026. What the publisher says, read 26 September 2026: the Department states no license for the files at ope.ed.gov/athletics, and the site carries no terms, copyright or privacy page of its own (its robots.txt redirects to www.ed.gov). The Department's machine-readable catalog, https://data.ed.gov/data.json, has one EADA entry, "Equity in Athletics Disclosure Act Survey, 2013", publisher "Office of Postsecondary Education (OPE)", "license": "http://www.opendefinition.org/licenses/cc-zero", "accessLevel": "public", and no entry for any other year. www.ed.gov's public-domain notice "does not pertain to information at web sites other than www.ed.gov". The condition: the Office of Postsecondary Education is named as the publisher on every page built from this source, the file is linked, and no seal or "official" framing is used.
Cadence
Annual. Colleges report each fall for the year before; the Department publishes the year's files the following spring (the 2024–25 zip's files are dated April 2026). A file is made once, at the close of collection, and "do[es] not include any corrections made following the close of the selected collection year".
This copy fetched
Rows held
34,843
Columns
53 columnssurvey_year, unitid, opeid, institution_name, city, state, zip, classification_code, classification_name, classification_other, sector_code, sector_name, enrollment_men, enrollment_women, enrollment_total, participants_men, participants_women, athletic_aid_men, athletic_aid_women, athletic_aid_coed, athletic_aid_total, recruiting_men, recruiting_women, recruiting_coed, recruiting_total, head_coach_salary_men, head_coach_salary_women, head_coach_salary_coed, head_coaches_men, head_coaches_women, head_coaches_coed, head_coach_salary_per_fte_men, head_coach_salary_per_fte_women, head_coach_salary_per_fte_coed, head_coach_fte_men, head_coach_fte_women, head_coach_fte_coed, assistant_coach_salary_men, assistant_coach_salary_women, assistant_coach_salary_coed, assistant_coaches_men, assistant_coaches_women, assistant_coaches_coed, assistant_coach_salary_per_fte_men, assistant_coach_salary_per_fte_women, assistant_coach_salary_per_fte_coed, assistant_coach_fte_men, assistant_coach_fte_women, assistant_coach_fte_coed, revenue_not_allocated, expense_not_allocated, grand_total_revenue, grand_total_expense
Pages built on it
Sports

What is committed here is a summary. The full table has 34,843 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is f00bc74902bcfefa7b6e466c777381f33a7e84b6ef3867b47f06ebb2a27291c7, so every figure here can be traced to the bytes it came from.

Every committed table for Equity in Athletics Disclosure Act data — every college athletics program, by institution, with its size and a download.
TableRowsSizeDownload
College athletics revenue, expenses, participants and aid by survey year, state and division7,571266 KBCSV
Every college reporting athletics, with its revenue, expenses and participants2,037123 KBCSV

What it cannot show

— 5 limitations
  • Coaches' salaries are averages, and no coach is named. A college's head-coach salary for men's teams is the average over all of its head coaches of men's teams — six at a college with six men's teams — in the Department's words, "all wages and bonuses an institution pays a coach as compensation attributable to coaching". It is not any one coach's pay, and pay from a booster club, a shoe contract or a media deal is not in it.
  • Revenue is not profit, and at most colleges it is set equal to expenses. The Department's revenue includes "state or other government support" and "student activity fees", so a college that covers its athletics budget from its own funds reports that money as revenue. In 2024–25, 1,288 of the 2,037 colleges reported revenues exactly equal to their expenses.
  • The figures are the colleges' own, as submitted, and are not audited by the Department. A year's file is made at the close of collection and "do[es] not include any corrections made following the close of the selected collection year".
  • Only colleges that report are here. The Department's words: the data "are submitted annually as required by the Equity in Athletics Disclosure Act (EADA) ... by all co-educational postsecondary institutions that receive Title IV funding ... and that have an intercollegiate athletics program". A college outside that, or one that did not report, is absent. Measured in 2024–25: the Air Force, Military, Naval and Coast Guard academies are not in the file; the Merchant Marine Academy is.
  • A blank is not a zero. The file writes a figure blank where the college has no team of that kind — most have no coed teams — and a zero where the amount is zero.

Equity in Athletics Disclosure Act data — every varsity team, by college and sport

Publisher
Office of Postsecondary Education, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works, recorded in our words and not the Department's; decided 26 September 2026. What the publisher says, read 26 September 2026: the Department states no license for the files at ope.ed.gov/athletics, and the site carries no terms, copyright or privacy page of its own (its robots.txt redirects to www.ed.gov). The Department's machine-readable catalog, https://data.ed.gov/data.json, has one EADA entry, "Equity in Athletics Disclosure Act Survey, 2013", publisher "Office of Postsecondary Education (OPE)", "license": "http://www.opendefinition.org/licenses/cc-zero", "accessLevel": "public", and no entry for any other year. www.ed.gov's public-domain notice "does not pertain to information at web sites other than www.ed.gov". The condition: the Office of Postsecondary Education is named as the publisher on every page built from this source, the file is linked, and no seal or "official" framing is used.
Cadence
Annual. Colleges report each fall for the year before; the Department publishes the year's files the following spring (the 2024–25 zip's files are dated April 2026). A file is made once, at the close of collection, and "do[es] not include any corrections made following the close of the selected collection year".
This copy fetched
Rows held
291,236
Columns
116 columnssurvey_year, unitid, sport_code, sport, sport_published, participants_men, participants_women, participants_coed_men, participants_coed_women, participants_men_total, participants_women_total, head_coaches_men_full_time_male, head_coaches_men_part_time_male, head_coaches_men_full_time_employee_male, head_coaches_men_part_time_employee_male, head_coaches_men_full_time_female, head_coaches_men_part_time_female, head_coaches_men_full_time_employee_female, head_coaches_men_part_time_employee_female, head_coaches_men, head_coaches_women_full_time_male, head_coaches_women_part_time_male, head_coaches_women_full_time_employee_male, head_coaches_women_part_time_employee_male, head_coaches_women_full_time_female, head_coaches_women_part_time_female, head_coaches_women_full_time_employee_female, head_coaches_women_part_time_employee_female, head_coaches_women, head_coaches_coed_full_time_male, head_coaches_coed_part_time_male, head_coaches_coed_full_time_employee_male, head_coaches_coed_part_time_employee_male, head_coaches_coed_full_time_female, head_coaches_coed_part_time_female, head_coaches_coed_full_time_employee_female, head_coaches_coed_part_time_employee_female, head_coaches_coed, head_coaches_all_full_time_male, head_coaches_all_part_time_male, head_coaches_all_full_time_employee_male, head_coaches_all_part_time_employee_male, head_coaches_all_full_time_female, head_coaches_all_part_time_female, head_coaches_all_full_time_employee_female, head_coaches_all_part_time_employee_female, head_coaches_all, assistant_coaches_men_full_time_male, assistant_coaches_men_part_time_male, assistant_coaches_men_full_time_employee_male, assistant_coaches_men_part_time_employee_male, assistant_coaches_men_full_time_female, assistant_coaches_men_part_time_female, assistant_coaches_men_full_time_employee_female, assistant_coaches_men_part_time_employee_female, assistant_coaches_men, assistant_coaches_women_full_time_male, assistant_coaches_women_part_time_male, assistant_coaches_women_full_time_employee_male, assistant_coaches_women_part_time_employee_male, assistant_coaches_women_full_time_female, assistant_coaches_women_part_time_female, assistant_coaches_women_full_time_employee_female, assistant_coaches_women_part_time_employee_female, assistant_coaches_women, assistant_coaches_coed_full_time_male, assistant_coaches_coed_part_time_male, assistant_coaches_coed_full_time_employee_male, assistant_coaches_coed_part_time_employee_male, assistant_coaches_coed_full_time_female, assistant_coaches_coed_part_time_female, assistant_coaches_coed_full_time_employee_female, assistant_coaches_coed_part_time_employee_female, assistant_coaches_coed, assistant_coaches_all_full_time_male, assistant_coaches_all_part_time_male, assistant_coaches_all_full_time_employee_male, assistant_coaches_all_part_time_employee_male, assistant_coaches_all_full_time_female, assistant_coaches_all_part_time_female, assistant_coaches_all_full_time_employee_female, assistant_coaches_all_part_time_employee_female, assistant_coaches_all, opexp_per_participant_men, opexp_per_team_men, opexp_per_participant_women, opexp_per_team_women, opexp_men_women, opexp_per_participant_coed_men, opexp_per_team_coed_men, opexp_per_participant_coed_women, opexp_per_team_coed_women, opexp_coed, opexp_per_participant_men_total, opexp_per_team_men_total, opexp_per_participant_women_total, opexp_per_team_women_total, opexp_total, revenue_men, revenue_women, revenue_men_women, revenue_coed_men, revenue_coed_women, revenue_coed, revenue_men_total, revenue_women_total, revenue_total, expense_men, expense_women, expense_men_women, expense_coed_men, expense_coed_women, expense_coed, expense_men_total, expense_women_total, expense_total
Pages built on it
Sports

What is committed here is a summary. The full table has 291,236 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is b4e538098fadc825c958ad768a2eb2f435075e5aaa2be3d10f69c05e91003fde, so every figure here can be traced to the bytes it came from.

Every committed table for Equity in Athletics Disclosure Act data — every varsity team, by college and sport, with its size and a download.
TableRowsSizeDownload
Every college team in the newest survey year, by college and sport18,045356 KBCSV
College teams, participants, revenue and expenses by sport and survey year62825 KBCSV

What it cannot show

— 7 limitations
  • Revenue is not profit. A team's revenue includes what the college allocates to it from its own funds and student fees, and at most colleges total revenue is set equal to total expenses (eada_institutions). A team whose revenue exceeds its expenses has not thereby made money.
  • The 2019–20 file also lists sports a college does not field: 44,556 rows, every figure zero or blank, beside 17,612 teams. They are dropped, and so is any such row in any year.
  • The 2020–21 year is the pandemic year: 14,381 teams at 1,844 colleges, against 17,612 at 2,073 the year before. Seasons were canceled; the file is what colleges reported.
  • Coaches are counted two ways, by assignment (full or part time to the team) and by employment (a full- or part-time employee of the college). Each split sums to the team's total; the two are never added together.
  • A coed team's men's and women's shares are the college's split, and in 11 rows of 2024–25 differ from the team's total by a dollar of rounding. Totals are read from their own columns.
  • The figures are the colleges' own, as submitted, and are not audited by the Department, and a year's file includes no correction made after its collection closed.
  • A blank is not a zero. A figure is blank where the college has no team of that kind.

EOIR immigration court outcomes by judge, court and fiscal year

Publisher
U.S. Department of Justice, Executive Office for Immigration Review
License
Unless otherwise indicated, information on Department of Justice websites is in the public domain and may be copied and distributed without permission.
Cadence
Monthly, republished in place on or about the first of the month
This copy fetched
Rows held
139,047
Columns
13 columnsrelease, judge_code, court_code, fiscal_year, fiscal_year_complete, proceedings_recorded, proceedings_pending, proceedings_completed, completions_decided, completions_in_absentia, asylum_granted, asylum_denied, asylum_neither

What is committed here is a summary. The full table has 139,047 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 376aadb27117ea31265ff8aeee41271f868de58b8a664201fb0440e4121f94c2, so every figure here can be traced to the bytes it came from.

Every committed table for EOIR immigration court outcomes by judge, court and fiscal year, with its size and a download.
TableRowsSizeDownload
Decisions by court and fiscal year3,28850 KBCSV
Asylum decisions, by judge, five complete years9319 KBCSV
Proceedings completed, by judge1,6859 KBCSV
Decisions nationally, by fiscal year752 KBCSV

What it cannot show

— 19 limitations
  • This table is an aggregate and the rows behind it are not kept. B_TblProceeding and tbl_Court_Appln are 33.2 million rows and 4.4 GB inflated; they are streamed, counted and discarded. The checks that need the per-proceeding detail run in the same pass and their results are in the snapshot's release.json. The bytes are recoverable: the whole 4.24 GiB release is archived every month.
  • A row is not a cohort and two of its columns must not be added together. proceedings_recorded and proceedings_pending are counted on the year EOIR recorded the proceeding (INPUT_DATE); proceedings_completed, completions_decided, completions_in_absentia and the three asylum columns are counted on the year it completed (COMP_DATE). A proceeding recorded in 2019 and completed in 2024 appears in both years and is one proceeding.
  • The asylum grant rate is granted / (granted + denied), and granted is five codes, not one: G a grant, I an in-court stipulated grant and P a paper stipulated grant, both added in 2021, plus C a conditional grant and F a full grant, both retired and both still in the historical rows. A rate computed as G / (G + D) leaves four of them out of the numerator and the denominator at once. Abandoned, withdrawn, transferred, administratively closed, reserved and not-adjudicated applications are asylum_neither and are in neither.
  • EOIR's own list of the decisions an asylum application may receive (strAllowedDecisionCodes on the ASYL row of tblLookUp_Appln) does not include C, and C is in the data. The classification is therefore checked against tblLookupCourtAppDecisions, which is every code and what each one means, and the ingest fails if EOIR adds a code or rewords one.
  • A grant rate over a handful of decisions is not a rate. A judge who decided thirty asylum applications moves several points on one of them, and any page publishing a rate states the count and declines the rate below a floor, the way the crime hub declines a rate for a jurisdiction that did not report.
  • A judge's mix of represented and unrepresented applicants is not the judge's doing and it moves the rate more than the judge does. The representation share is in tbl_RepsAssigned, which is not loaded here, and no page publishes a grant rate without it.
  • The current fiscal year is incomplete and fiscal_year_complete says so. The release is generated on or about the first of a month, so the fiscal year it falls in is missing everything after that date. A total for an incomplete year is not comparable with a total for a complete one.
  • proceedings_pending counts proceedings with no completion date as of the release, attributed to the fiscal year they were recorded in. It is a snapshot of the backlog by age, not a measure of anything that happened in that year, and it changes every month for every past year.
  • INPUT_DATE is the date EOIR recorded the proceeding in its system, which is not the date the charging document was issued (OSC_DATE) and not the date a case was filed. proceedings_recorded is named for what the column is.
  • The proceeding decision breakdown is not here. B_TblProceeding.DEC_CODE does not decode on its own: tblLookupCourtDecision keys on case type and decision type as well, and under AOC the code C is "Conditional Grant" with one decision type and "Change of Venue" with the other. completions_decided counts completions that carry any decision code; what the decision was is deferred to a second pass.
  • 476,090 proceedings have no usable INPUT_DATE, 2.8% of the release. They are in no recorded year, so they are absent from proceedings_recorded and from proceedings_pending; 12,722 of them are pending, which is backlog this table cannot date. Both counts are in the snapshot's release.json under rows_set_aside.
  • 2,209 asylum applications granted or denied in the release are not in this table, 0.16% of the 1,373,075: the proceeding they were decided in has no completion date, or is one of the fourteen in a test base city. They are in the census in release.json, because the table dates an asylum decision by the completion of the proceeding it was decided in.
  • 1,363 proceedings carry an INPUT_DATE of 1 January 1900. That is a placeholder for an unknown date, not a year, and it is read as no date: 1,356 of those proceedings completed in a later year and are counted there, and the 7 still pending are backlog this table cannot date. Left alone it would have been a fiscal year of its own, 51 years before the next one in the file.
  • One proceeding can carry several asylum applications — a lead case and its riders are separate rows against the same IDNPROCEEDING — so the asylum columns count applications and can exceed proceedings_completed in the same row. A grant rate is over applications. A page that calls them cases is wrong by however many riders there were.
  • ASYW, asylum withholding, is a separate application code and is not counted here. It is nearly as common as ASYL — 589 against 673 in a sample of one release — and is granted on a different standard, so any filter that takes both is measuring two things at once.
  • judge_code is also null where EOIR recorded AAA, its "<All Judges>" sentinel: 64 judge-court-year groups and 93 completed proceedings in this release. The proceedings are real and stay in every total; the judge is not a person and eoir_judges excludes the same code from the roster, so a page built on this column cannot make one.
  • Proceedings in EOIR's own test base cities (OIT, OIN) are dropped and counted, not published. OIT carries 13 of them in this release, eleven completed, in fiscal 2022 to 2026. They are test records in production data and they are not proceedings. The count is in release.json under rows_set_aside.
  • judge_code and court_code are null where EOIR recorded the proceeding without one. Those rows are kept, because they still belong in the court's and the judiciary's totals, and dropping them would make every total quietly smaller with nothing red anywhere.
  • Certain data in the wider release — custody, detention and date of entry — is not sourced from EOIR. None of it is in this table. EOIR's Readme.txt says: "For the most accurate records on this information, please contact the Department of Homeland Security."

EOIR immigration judges and courts

Publisher
U.S. Department of Justice, Executive Office for Immigration Review
License
Unless otherwise indicated, information on Department of Justice websites is in the public domain and may be copied and distributed without permission.
Cadence
Monthly, republished in place on or about the first of the month
This copy fetched
Rows held
12,370
Columns
13 columnsrelease, judge_code, judge_name, judge_active, judge_is_person, judge_record_created, court_code, court_name, court_city, court_state, court_region, court_active, assignment_active
Download
Immigration judges and their court assignments — gzipped CSV, 87 KB

What is committed here is a summary. The full table has 12,370 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 817fdee640da11ce47d8add10b3dd2305d14bf3c0959ddcffc0f4deb04bf272b, so every figure here can be traced to the bytes it came from.

What it cannot show

— 9 limitations
  • The roster is not a list of judges. 88 of the 1,569 entries EOIR marks active are docket slots rather than people, and judge_is_person separates them. A count of rows, or of active rows, is a count of roster entries — which is what 1,784 and 1,569 are, and what the immigration hub plan briefly published as a judge count.
  • This table holds no decisions. It is a roster and an assignment history, and a page built on it alone would carry a judge's name, their courts and nothing a reader came for. Case outcomes are in B_TblProceeding and tbl_Court_Appln, which are 4.4 GB compressed and are not loaded here.
  • datCreatedOn is the date EOIR created the database record, not the date a judge was appointed. 1,388 of the 1,785 roster rows carry the same August 10, 2003 timestamp, which is the day the table was migrated. It must never be rendered as an appointment date.
  • A judge is assigned to as many as 72 courts, so a judge has no single court and any page that names one is choosing. The active flag on an assignment says whether EOIR still records it, not where the judge sits today.
  • The 97 base cities are not 97 immigration courts. They include the Board of Immigration Appeals (BIA), the Office of the Chief Immigration Judge (CIJ), EOIR's Office of Management and Administration (HDQ) and two test rows. Only the test rows and the ALL sentinel are excluded here; the three non-court units are in this table, with judges assigned to each, and a consumer that publishes a page per row has to drop them itself. This limitation said they were "excluded here by code" for two days and they never were — corrected 21 September 2026 after a slug collision in the web build, not after anyone re-read the sentence.
  • A city and state do not identify a court. Eleven city-state pairs carry more than one: New York has four base cities, Los Angeles four, Houston four. BASE_CITY_NAME is "IMMIGRATION COURT" for most rows and so does not separate them either. The base-city code is the key.
  • Forty assignment rows, 0.32% of the table, name 26 judge codes absent from the roster and one court code absent from the base-city table. They are dropped, named in the log, and the ingest fails if the dropped share ever exceeds 1%.
  • EOIR overwrites the release URL in place and every dated variant 404s, so a release that is not archived in the month it is published cannot be recovered. This source reads three members by range request and does not archive the 4.24 GiB release; a separate daily job does.
  • Certain data in the wider release — custody, detention and date of entry — is not sourced from EOIR. EOIR's own Readme.txt says: "For the most accurate records on this information, please contact the Department of Homeland Security." None of those fields are in this table, and any later table carrying them is attributed to DHS-via-EOIR rather than to EOIR.

FBI Crime Data Explorer — agencies and the population each serves

Publisher
Federal Bureau of Investigation, Criminal Justice Information Services Division
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual, following the UCR release
This copy fetched
Rows held
19,636
Columns
10 columnsori, state_abbr, agency_name, agency_type, county, latitude, longitude, is_nibrs, population_year, population
Download
Police agencies and the populations they serve — gzipped CSV, 410 KB

What it cannot show

— 5 limitations
  • The roster carries no population. The figure here comes from the population block inside a crime call for each agency, which is one request per agency: there is no bulk endpoint, and participation/…, summarized/agencies/{ST}/… and agency/byStateAbbr/{ST}/population were each probed and answer 404.
  • The population is the FBI's own attribution for the agency, not a Census figure for a place. Two agencies can serve overlapping populations — a city police department and the county sheriff around it — so agency populations must never be summed to a county or state total.
  • An agency with no population is one the FBI published none for, usually because it reported no crime data in the year asked for. It is carried with a null rather than a zero, because a zero would divide into a rate.
  • Reporting is voluntary. An agency's presence in this list says the FBI holds a record for it, not that it submitted anything.
  • The population belongs to one year. A rate built from it and a count from another year is a rate of two different things, so the year is carried in a column and shown on the page rather than assumed.

FBI Crime Data Explorer — crime reported by each agency

Publisher
Federal Bureau of Investigation, Criminal Justice Information Services Division
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Quarterly, following the UCR release
This copy fetched
Rows held
195,276
Columns
5 columnsori, offense, year, measure, value
Download
Offenses reported by each agency, by year — gzipped CSV, 1.4 MB

What it cannot show

— 6 limitations
  • The population is the FBI's figure for that agency in that year, and it changes annually for every agency measured — up to 11.4%. A rate for a year divides by that year's population; the single figure in fbi_agencies is for the 25,000 threshold and is never a denominator for a historical year. Using it would be wrong by a median 4.6% and up to 93.4%.
  • Agency populations overlap: a city police department and the sheriff around it both serve some of the same residents and both report. Summed across a state they come to 101% of its covered population at the median and 128% in Louisiana, so these counts are never added to a county, state or national total.
  • Only agencies serving 25,000 or more are crawled, and only City and County agencies. That is measured: of 11,719 agencies with a published population, 33% of County and 17% of City clear the threshold, against 1% of State Police and zero across Other, Other State Agency, University or College and Tribal. An agency absent from this table is one with no page, not one that reported nothing.
  • Reporting is voluntary and monthly. A month with no submission is null, not zero, and a year assembled from a partial set of months is a partial year. The page states how many months it holds.
  • The population is the FBI's own attribution for that agency in that month, and it is the population the agency serves — so unlike a state figure it needs no coverage adjustment. Our arithmetic agrees with the FBI's published rate to within 0.015%, which is its rounding to two decimals.
  • An offense is one reported to and recorded by that agency. It is not a conviction and does not include what was never reported.

FBI Crime Data Explorer — Arrests

Publisher
Federal Bureau of Investigation, Criminal Justice Information Services Division
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual release with quarterly updates
This copy fetched
Rows held
45,500
Columns
6 columnsscope, state_abbr, year, dimension, category, arrests
Download
Arrests by offense, age, sex and race — gzipped CSV, 225 KB

What it cannot show

— 6 limitations
  • Arrests are events, not people. One person arrested three times in a year is three arrests, so an arrest count is never a count of individuals and a rate built from it is never a share of a population that was arrested.
  • Race is recorded by the arresting agency, not reported by the person arrested. The FBI's categories are not the Census Bureau's: the FBI records one race per arrestee and its "Multiple" category is an officer's classification, where the Census "Two or more races" figure is self-identified. The two are not the same population.
  • There is no ethnicity field in this data at all — not an empty one, absent. No Hispanic or Latino arrest figure can be derived from it.
  • Reporting is voluntary and coverage varies by agency and by year, so a change in an arrest count can be a change in which agencies reported.
  • The offense vocabularies overlap. "Offense Name", "Offense Category" and "Offense Breakdown" are three different groupings of the same arrests and must never be summed together or across each other.
  • Rape appears under both the current and the legacy definition, which the FBI revised in 2013. The two are not the same offense and are not added together.

FBI Crime Data Explorer API

Publisher
Federal Bureau of Investigation, Criminal Justice Information Services Division
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual release with quarterly updates
This copy fetched
Rows held
832,124
Columns
7 columnsscope, state_abbr, offense, year, month, measure, value
Pages built on it
Crime and justice
Download
Offenses and rates by state and year — gzipped CSV, 4.0 MB

What it cannot show

— 5 limitations
  • Reporting to the FBI is voluntary. Coverage varies by agency and by year, so a change in a state total can be a change in which agencies reported rather than a change in offenses.
  • Counts are offenses reported to law enforcement. They are not convictions and do not measure offenses that were never reported.
  • Rates use the FBI's own population denominators, which this pipeline stores alongside the counts so a published rate can be recomputed exactly as the FBI computed it.
  • The UCR program moved from summary reporting to NIBRS, and agencies transitioned in different years. A change in a state's total can be a change in who reported, which is why agency participation is stored with the crime figures.
  • Agencies revise submissions. A figure is only as current as the snapshot it was drawn from.

FBI Crime Data Explorer — Hate Crime

Publisher
Federal Bureau of Investigation, Criminal Justice Information Services Division
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual release with quarterly updates
This copy fetched
Rows held
106,964
Columns
7 columnsscope, state_abbr, year, section, dimension, category, incidents
Download
Hate crime incidents by bias motivation — gzipped CSV, 379 KB

What it cannot show

— 6 limitations
  • Reporting hate crime to the FBI is voluntary, and more unevenly done than reporting offenses generally. A jurisdiction reporting zero hate crimes may have had none, or may not have submitted; these figures cannot tell the two apart, so a count of zero is never presented as an absence of hate crime.
  • An incident is counted where an agency judged that a bias motivated it. That judgment is the agency's, made under its own reading of the FBI's definitions, and it is not a court's finding.
  • Counts are small. A change of a handful of incidents moves a jurisdiction's total by a large percentage, and a percentage change on a small base is not evidence of a trend.
  • The bias, offense, location and offender breakdowns each describe the same incidents from a different angle and must never be summed together or across each other.
  • An incident may carry more than one bias, which the FBI records as "Multiple Bias" rather than counting it under each.
  • Offender race and ethnicity are recorded by the reporting agency where known, and are unknown for a substantial share of incidents. That share is published alongside them.

FBI Crime Data Explorer — Police Employment and LEOKA

Publisher
Federal Bureau of Investigation, Criminal Justice Information Services Division
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual release with quarterly updates
This copy fetched
Rows held
2,362
Columns
6 columnscollection, year, dimension, series, category, value
Download
Police employment by agency and year — gzipped CSV, 12 KB

What it cannot show

— 5 limitations
  • Police employment is available from this API at national scope only. The state endpoint answers with every field null, confirmed across several states and year ranges, and passing a state as a query parameter on the national path is ignored and returns national figures. There is therefore no state employment page, and the absence is a limit of the source rather than of this pipeline.
  • Employment counts are of staff at agencies that reported. A change can be a change in which agencies reported rather than in how many officers there are, and the population those agencies served is published alongside so a per-capita figure can be read against the right denominator.
  • An officer assaulted is an assault reported by the officer's agency, not a charge or a conviction. The circumstance categories are the agency's classification.
  • LEOKA counts are small, particularly officers killed. A change of a few moves a percentage a long way, and a percentage change on a small base is not a trend.
  • The LEOKA breakdowns each describe the same incidents from a different angle — weapon, location, assignment, lighting — and are never summed together.

FBI Crime Data Explorer — Supplementary Homicide Reports

Publisher
Federal Bureau of Investigation, Criminal Justice Information Services Division
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Annual release with quarterly updates
This copy fetched
Rows held
288,288
Columns
7 columnsscope, state_abbr, year, subject, dimension, category, count
Download
Homicide victims and offenders — gzipped CSV, 1.1 MB

What it cannot show

— 6 limitations
  • This is homicide alone. The Supplementary Homicide Reports describe murder and non-negligent manslaughter and nothing else, so no figure here is about crime victims generally, and none of it extends to any other offense.
  • The breakdowns are marginals, not a cross-tabulation. Victim age, victim sex and victim race each count the same victims a different way and each sums to the same total, so a figure combining two of them — victims of one race in one age band — is not in this data and cannot be derived from it.
  • Age, sex, race, ethnicity, relationship, weapon and circumstance are what the reporting agency recorded, under its own reading of the FBI's definitions. Where an agency did not record a value it is counted as unknown, and the unknown share is published beside every breakdown rather than removed from it.
  • Reporting is voluntary and not every agency submits a supplementary report for every homicide it records. The totals here are therefore lower than the homicide counts in the summarized collection, which is the FBI's own count of the offense, and the two are not interchangeable.
  • An offender's race or relationship is only known where the offender is known. A homicide that was never cleared contributes a victim but an unknown offender, so offender breakdowns describe solved homicides more than they describe all of them.
  • Three race categories are published and have never carried a value in any year or jurisdiction — "Multiple", "Not Specified", and "Asian, Native Hawaiian, or Other Pacific Islander". They are the publisher's schema rather than a finding of zero, and are not rendered as zeros.

FDIC failures — every bank failure and assistance transaction since 1934

Publisher
Federal Deposit Insurance Corporation
License
U.S. Government work. This is our record, not the FDIC's words: Kian decided on September 26, 2026 to record BankFind as NCES and IPEDS are recorded — the FDIC is an independent agency of the United States, and BankFind is its own record of the institutions it insures and their offices and financial reports, U.S. Government works under 17 U.S.C. 105. The publisher's own label, from its catalog https://www.fdic.gov/data.json (DCAT-US 3.0, modified September 23, 2026, read September 26, 2026), on the distribution "Bank Failures and Assistance Data API" (/failures) and every other FDIC distribution: "license": "https://www.usa.gov/publicdomain/label/1.0/", "accessRights": "Public", use restriction "Unrestricted", and "This distribution is publicly accessible. It may be copied, published, distributed, transmitted, cited, and adapted consistent with the stated license." The FDIC's website policies (https://www.fdic.gov/about/website-policies, last updated May 12, 2026) state no copyright position of the FDIC's own and disclaim warranty: "the FDIC makes no express or implied warranty regarding such information or data". The conditions, in this site's register of licenses: the FDIC is named as publisher with the endpoint linked and the record date, no FDIC seal, and nothing "official".
Cadence
As banks close. The FDIC's failed bank list carries a closing within a day (Nano Banc, closed September 25, 2026, was on it on 26 September); the failures index behind the API is rebuilt now and then, and the one serving on September 26, 2026 was built August 25, 2026, 14:13 UTC, with its newest failure on 17 July
This copy fetched
Rows held
4,119
Columns
20 columnsid, cert, fin, name, city, state, failure_date, resolution_date, resolution, transaction_type, charter_class, insurance_fund, total_assets, total_deposits, estimated_loss, estimated_loss_as_of, acquirer_name, acquirer_city, acquirer_state, listed_by
Pages built on it
Banks and credit unions

What is committed here is a summary. The full table has 4,119 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is a80d608fae279c56d7c99ec9b714d7d92da0bf6172d8ef31b1f56b00f4ec94d4, so every figure here can be traced to the bytes it came from.

Every committed table for FDIC failures — every bank failure and assistance transaction since 1934, with its size and a download.
TableRowsSizeDownload
Bank failures and assistance transactions by year since 1934, with assets and estimated losses1071 KBCSV
Every bank failure of the two newest years with one, by name81 KBCSV

What it cannot show

— 6 limitations
  • The estimated loss is an estimate, and changes. It is the FDIC's estimate of the cost to the insurance fund as of the date beside it, revised as a receivership is wound down; it is blank on 637 records (539 before 1977) and negative on 12, an estimated gain.
  • A loss of 0 is the FDIC's figure, not an explanation. 73 records carry an estimated loss of exactly 0, among them Signature Bank (12 March 2023, $110.4 billion of assets). The file does not say why; a page prints the FDIC's figure with that said, and nothing else.
  • Assets and deposits are from the last report before the failure, not the day of it; the FDIC notes they may differ from the institution's last report filed.
  • The newest failures may be missing. The index is rebuilt now and then, not at each closing; the failed bank list is current, and the freshness check compares the two.
  • Before 1977 many records have no certificate number (488, from 1934 to 1976) and none of those has a loss estimate, so they cannot be linked to an institution's history.
  • Dollars are each year's own, not adjusted for inflation.

FDIC BankFind — institutions (every charter the FDIC has insured, active and inactive)

Publisher
Federal Deposit Insurance Corporation
License
U.S. Government work. This is our record, not the FDIC's words: Kian decided on September 26, 2026 to record BankFind as NCES and IPEDS are recorded — the FDIC is an independent agency of the United States, and BankFind is its own record of the institutions it insures, U.S. Government works under 17 U.S.C. 105. The publisher's own label, from its catalog https://www.fdic.gov/data.json (DCAT-US 3.0, modified September 23, 2026, read September 26, 2026), on the distribution "FDIC Banking Demographic and Headquarter Locations" and every other FDIC distribution: "license": "https://www.usa.gov/publicdomain/label/1.0/", "accessRights": "Public", use restriction "Unrestricted", and "This distribution is publicly accessible. It may be copied, published, distributed, transmitted, cited, and adapted consistent with the stated license." The FDIC's website policies (https://www.fdic.gov/about/website-policies, last updated May 12, 2026) state no copyright position of the FDIC's own and disclaim warranty: "the FDIC makes no express or implied warranty regarding such information or data". The conditions, in this site's register of licenses: the FDIC is named as publisher with the file linked and its run date, no FDIC seal, nothing "official", and "FDIC-insured" printed as the FDIC's field, never as an assurance of ours.
Cadence
Weekly. The FDIC: "Institutions and Locations endpoints are updated weekly and limited to the current data" (the Bulk Data Download page). The file of September 25, 2026 was stamped RUNDATE 09/25/2026 and last modified 25 Sep 2026 12:31 GMT
This copy fetched
Rows held
27,835
Columns
34 columnscert, name, active, fdic_insured, address, city, state_abbr, state_name, county_fips, county_name, zip, bank_class, charter_agency, primary_regulator, fed_rssd, established_date, insured_date, end_date, successor_cert, holding_company, holding_company_rssd, website, community_bank, mutual, specialization, offices, offices_domestic, offices_foreign, report_date, total_assets, total_deposits, domestic_deposits, equity_capital, run_date
Pages built on it
Banks and credit unions

What is committed here is a summary. The full table has 27,835 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 77c7ab6d4ad945579ea098c29079e920e2e15916b525859cee52c430cb208802, so every figure here can be traced to the bytes it came from.

Every committed table for FDIC BankFind — institutions (every charter the FDIC has insured, active and inactive), with its size and a download.
TableRowsSizeDownload
FDIC-insured banks by headquarters state, with community banks, offices, assets and deposits552 KBCSV
The 100 largest banks by assets and by deposits, each ranked among every bank that reported1034 KBCSV

What it cannot show

— 5 limitations
  • The financial figures are as of each charter's newest report, not one date. For the 4,220 active institutions that filed for 30 June 2026 they are that quarter's; a charter that has not filed yet, or closed long ago, carries its last report (its report date says which), and 4 active institutions have none at all, including one opened on September 14, 2026. Adding up total assets across the file mixes years unless the report date is held fixed.
  • Assets and deposits are as reported on the Call Report, consolidated. A bank that owns another bank reports it inside its own figures, and a holding company's non-bank parts are not in any row. The FDIC's total for the industry (its Quarterly Banking Profile) counts a slightly different set of charters: 4,238 at 30 June 2026 against the 4,231 active rows here.
  • Inactive is not failed. 23,603 of the 27,834 charters are inactive: merged, acquired, converted, voluntarily closed or failed. This file does not say which; the failures and structure-change sources do. fdic_insured is 1 on every row, open or closed, so whether a charter is insured today is active, which the FDIC defines as "currently open and insured by the FDIC".
  • Headquarters only. The address, city, county and state are the main office's. A bank's branches are in the locations source, and a bank headquartered in one state can hold most of its deposits in others.
  • No income or return figures. The file writes zero where the FDIC has no figure for net income and return on equity, so those are not read here; they come from the FDIC's quarterly financials.

FDIC BankFind — locations (every current office of an FDIC-insured institution)

Publisher
Federal Deposit Insurance Corporation
License
U.S. Government work. This is our record, not the FDIC's words: Kian decided on September 26, 2026 to record BankFind as NCES and IPEDS are recorded — the FDIC is an independent agency of the United States, and BankFind is its own record of the institutions it insures and their offices, U.S. Government works under 17 U.S.C. 105. The publisher's own label, from its catalog https://www.fdic.gov/data.json (DCAT-US 3.0, modified September 23, 2026, read September 26, 2026), on the distribution "FDIC Bank Locations" and every other FDIC distribution: "license": "https://www.usa.gov/publicdomain/label/1.0/", "accessRights": "Public", use restriction "Unrestricted", and "This distribution is publicly accessible. It may be copied, published, distributed, transmitted, cited, and adapted consistent with the stated license." The FDIC's website policies (https://www.fdic.gov/about/website-policies, last updated May 12, 2026) state no copyright position of the FDIC's own and disclaim warranty: "the FDIC makes no express or implied warranty regarding such information or data". The conditions, in this site's register of licenses: the FDIC is named as publisher with the endpoint linked and its index date, no FDIC seal, and nothing "official".
Cadence
Weekly. The FDIC: "Institutions and Locations endpoints are updated weekly and limited to the current data" (the Bulk Data Download page). The index of September 25, 2026 was built at 11:43 UTC with RUNDATE 09/25/2026
This copy fetched
Rows held
78,077
Columns
20 columnsuninum, cert, office_number, office_name, main_office, service_type, address, address_2, city, county_name, county_fips, state_abbr, state_name, zip, latitude, longitude, cbsa_code, cbsa_name, established_date, run_date
Pages built on it
Banks and credit unions

What is committed here is a summary. The full table has 78,077 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is e13adf12d6fe4a77843b38767639b5768325182c627caf8993c116ff3eb06b57, so every figure here can be traced to the bytes it came from.

Every committed table for FDIC BankFind — locations (every current office of an FDIC-insured institution), with its size and a download.
TableRowsSizeDownload
Offices abroad, by the bank that holds them240 KBCSV
Offices of FDIC-insured banks by state and service type4853 KBCSV

What it cannot show

— 4 limitations
  • Current offices only. The endpoint holds the offices open on its run date; an office that closed last month is not here, and neither is any history. Openings and closings are in the FDIC's structure-change history, and past years' offices in the Summary of Deposits.
  • An office is not a branch with deposits. 78,082 offices include 4,228 main offices, loan production offices, drive-throughs, administrative offices, trust offices and mobile offices; the service type says which. Deposits per office are in the Summary of Deposits, as of each 30 June, not here.
  • 802 offices are abroad, all of them branches of U.S. banks (Citibank, Bank of America, JPMorgan Chase and others) in cities such as London, Nassau and Buenos Aires. They have no state, county or coordinates, and no country field: the city is all the FDIC gives.
  • Coordinates in the Pacific territories are unreliable. In the fifty states and DC every office has coordinates inside the country. In Micronesia, the Marshall Islands and Palau several point to the wrong ocean or have latitude and longitude swapped (one reads latitude 138), and one office in the Northern Marianas has none. Latitudes beyond ±90 are null here; the others are kept as published.

FDIC Summary of Deposits — every office's deposits at June 30, 1994 to the newest survey

Publisher
Federal Deposit Insurance Corporation
License
U.S. Government work. This is our record, not the FDIC's words: Kian decided on September 26, 2026 to record BankFind as NCES and IPEDS are recorded — the FDIC is an independent agency of the United States, and BankFind is its own record of the institutions it insures and their offices' deposits, U.S. Government works under 17 U.S.C. 105. The publisher's own label, from its catalog https://www.fdic.gov/data.json (DCAT-US 3.0, modified September 23, 2026, read September 26, 2026), on the distribution "Branch Office Deposits" (/sod) and every other FDIC distribution: "license": "https://www.usa.gov/publicdomain/label/1.0/", "accessRights": "Public", use restriction "Unrestricted", and "This distribution is publicly accessible. It may be copied, published, distributed, transmitted, cited, and adapted consistent with the stated license." The FDIC's website policies (https://www.fdic.gov/about/website-policies, last updated May 12, 2026) state no copyright position of the FDIC's own and disclaim warranty: "the FDIC makes no express or implied warranty regarding such information or data". The conditions, in this site's register of licenses: the FDIC is named as publisher with the endpoint linked and the survey date, no FDIC seal, and nothing "official".
Cadence
Annual, as of 30 June, released in September. The FDIC on September 18, 2026: "The FDIC's Summary of Deposits (SOD) provides branch-level deposit information for more than 75,000 domestic offices operated by more than 4,200 FDIC-insured institutions ... The SOD includes historical data going back to 1994." The index serving the 2026 survey was built September 18, 2026, 13:38 UTC
This copy fetched
Rows held
2,898,836
Columns
17 columnsuninum, year, cert, fed_rssd, institution_name, branch_number, office_name, main_office, service_type, address, city, county_name, county_fips, state_abbr, state_name, zip, deposits
Pages built on it
Banks and credit unions

What is committed here is a summary. The full table has 2,898,836 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 14632c058fc9f09eea79c3fa79e967a76808b67835240be5aac204ef0ee10452, so every figure here can be traced to the bytes it came from.

Every committed table for FDIC Summary of Deposits — every office's deposits at June 30, 1994 to the newest survey, with its size and a download.
TableRowsSizeDownload
Bank offices and their deposits by county, at June 30 of the newest year3,19741 KBCSV
Bank offices and their deposits by state, at June 30 of every year from 19941,94427 KBCSV

What it cannot show

— 8 limitations
  • Deposits at one date a year. Each survey is the office's domestic deposits on 30 June of its year, as the institution reported them; nothing here says what an office held in any other month, and a change between two surveys is two snapshots, not a flow.
  • Deposits are booked where the institution books them, not where the depositor lives. An institution assigns each deposit to an office by its own rules, so an online bank or a headquarters office can hold deposits from customers across the country, and a county's total is what institutions booked there.
  • Some offices take no deposits. Drive-throughs, administrative, trust and mobile offices report zero: 2,189 of 75,859 offices in 2026. Zero is the FDIC's figure, not a missing one.
  • Domestic offices only, from 1994. Offices abroad are not surveyed. The surveys from 1987 to 1993 are published separately, in another layout, and are not in this source.
  • An office keeps its number when its institution changes. The row names the institution that held the office in that year's survey, so one office's history can cross a merger and show two banks' names.
  • Dollars are each year's own, not adjusted for inflation.
  • Connecticut's counties change in 2023. Through the 2022 survey its offices carry the eight old counties (09001 to 09015); from 2023 the nine planning regions the Census Bureau adopted (09110 to 09190). A Connecticut county's series stops in 2022 and a region's starts in 2023; neither is the other continued.
  • Older offices may have no unique number. 120,261 rows from 1994 to 2010 have no UNINUMBR, so those years' offices cannot all be followed to a current office; the survey identifies them by institution and office number within the year.

FDIC summary — insured institutions, offices, assets and deposits by state and year since 1934

Publisher
Federal Deposit Insurance Corporation
License
U.S. Government work. This is our record, not the FDIC's words: Kian decided on September 26, 2026 to record BankFind as NCES and IPEDS are recorded — the FDIC is an independent agency of the United States, and BankFind is its own record of the institutions it insures and their offices and financial reports, U.S. Government works under 17 U.S.C. 105. The publisher's own label, from its catalog https://www.fdic.gov/data.json (DCAT-US 3.0, modified September 23, 2026, read September 26, 2026), on the summary distribution under "Historical Bank Data" and every other FDIC distribution: "license": "https://www.usa.gov/publicdomain/label/1.0/", "accessRights": "Public", use restriction "Unrestricted", and "This distribution is publicly accessible. It may be copied, published, distributed, transmitted, cited, and adapted consistent with the stated license." The FDIC's website policies (https://www.fdic.gov/about/website-policies, last updated May 12, 2026) state no copyright position of the FDIC's own and disclaim warranty: "the FDIC makes no express or implied warranty regarding such information or data". The conditions, in this site's register of licenses: the FDIC is named as publisher with the endpoint linked and the year-end date, no FDIC seal, and nothing "official".
Cadence
Annual, as of 31 December. The FDIC rebuilds this index from the year-end Call Reports: the index serving on September 27, 2026 was built September 16, 2026, 13:42 UTC, and its newest year is 2025
This copy fetched
Rows held
7,761
Columns
10 columnsyear, charter_group, level, state, state_name, state_fips, institutions, offices, total_assets, total_deposits

What is committed here is a summary. The full table has 7,761 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 5e6fbe02c8eb9a1d92c2d815ba8162cf43a11984e351f2f82fbcfbda24ad3408, so every figure here can be traced to the bytes it came from.

Every committed table for FDIC summary — insured institutions, offices, assets and deposits by state and year since 1934, with its size and a download.
TableRowsSizeDownload
FDIC-insured institutions, offices, assets and deposits by state and year, by charter group7,627108 KBCSV
FDIC-insured institutions, offices, assets and deposits by year since 1934, by charter group1343 KBCSV

What it cannot show

— 5 limitations
  • Savings institutions start in 1984. Before it the file counts commercial banks only, so a count of all institutions steps up by 3,418 in 1984, when savings institutions join the series, not because that many opened.
  • A state's figures before 1984 are rounded and incomplete. A state's assets and deposits are rounded to the million on most rows before 1984 and the states fall short of the national figure by up to 15% (1977); from 1984 they add up to it exactly. The national figure is read from the national row.
  • Institutions are counted where they are headquartered, offices where they stand. A state can have offices and no institution, as Arizona's savings institutions did in 2019 to 2021 and three Pacific territories' banks do every year; offices of savings institutions are published by state only from 2019.
  • The count is the FDIC's historical-statistics count, which differs slightly from the Quarterly Banking Profile's. For 2019 to 2025 the commercial-bank count differs by up to 14 a year (2025: 3,806 here, 3,815 in the QBP); the savings count is the same.
  • Dollars are each year's own, not adjusted for inflation, and assets are consolidated, so a bank's foreign offices are counted in the state of its head office.

FEMA's FIFA World Cup Grant Program — every award under assistance listing 97.160

Publisher
U.S. Department of the Treasury, Bureau of the Fiscal Service
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
As FEMA reports. The awards were made in March and April 2026; the amounts paid out grow as the states draw the money down, which USAspending reports monthly.
This copy fetched
Rows held
9
Columns
15 columnsaward_id, usaspending_award_id, assistance_listing, recipient_name, recipient_uei, state_code, award_amount, total_outlays, description, awarding_agency, awarding_sub_agency, start_date, end_date, base_obligation_date, last_modified_date
Pages built on it
Sports
Download
Every FEMA World Cup security grant, by state — gzipped CSV, 1 KB

What is committed here is a summary. The full table has 9 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 0fc6f86830fb07fce9b82e7871394f8d72196a066e7559ca3c64389785974586, so every figure here can be traced to the bytes it came from.

What it cannot show

— 4 limitations
  • Obligated is not spent. The award amount is what FEMA obligated to each state. Total outlays is what has been paid out of it so far; on 26 September 2026 three of nine states had drawn any money. Neither is what the state spent on the ground.
  • The listing, not a keyword. Only awards under assistance listing 97.160 are here. World Cup costs paid through other programs, such as transit formula funds under listing 20.507, are not, because nothing in the record marks them as the World Cup's.
  • A state's recipient is its lead agency. Each grant goes to one state agency, which may pass money on to cities and host committees. Those subawards are not here.
  • One event. The 2028 Olympics' $1,000,000,000, from the same section, has no listing and no award yet.

Farm program payment files

Publisher
U.S. Department of Agriculture, Farm Service Agency
License
U.S. Government work. USDA states no license on the payment-files page itself; its site-wide Digital Rights and Copyright notice reads "Most information presented on the USDA Web site is considered public domain information. Public domain information may be freely distributed or copied, but use of appropriate byline/photo/image credits is requested. Attribution may be cited as follows: 'U.S. Department of Agriculture.'" It carves out third-party material — "Some materials on the USDA Web site are protected by copyright, trademark, or patent … USDA has made every attempt to identify and clearly label them" — which does not reach these files: they are FSA's own administrative record of disbursements it made, and are not labeled otherwise. The attribution is honoured rather than treated as optional: every page naming a figure from this file names the publisher and links the file. Read 22 September 2026 from a December 2024 archive snapshot, because usda.gov/policies-and-links returns 403 to scripted agents.
Cadence
Annual
This copy fetched
Rows held
77,404,426
Columns
18 columnspayment_year, state_fsa_code, state_fsa_name, county_fsa_code, county_fsa_name, payee_name, address_line, delivery_address, city, state_abbreviation, zip_code, disbursement_amount, payment_date, program_code, program_description, program_year, payment_at, ordinal
Pages built on it
Farm subsidies

What is committed here is a summary. The full table has 77,404,426 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from. The whole table is also a bulk download, below, as CSV and Parquet.

The full table’s SHA-256 is 2e5ed487f6f4d597a27b22d012d88871fcf86e0ed7e6fc5c5d472bbfd2259c8d, so every figure here can be traced to the bytes it came from.

Bulk download: Farm program payments

Every row, 2004–2025, built September 26, 2026. Our file is licensed CC BY 4.0: credit “American Factbook (americanfactbook.org)”. The source is the publisher’s own files, u.S. Government work; USDA's notice calls its site's information public domain.

Farm program payments, every disbursement. One row per disbursement in FSA's payment files, 2004 onward, with the street line of the mailing address left out.

  • farm-payments.csv.gz — gzip-compressed CSV, 1.14 GB, 77,404,426 rows
    SHA-256 c9de9e05866e6af33993a44bb7ae7a471354dc554d58fbed18b13991f90017ef
  • farm-payments.parquet — Parquet, 1.02 GB, 77,404,426 rows
    SHA-256 f704125cf62439b9ae9243db145fa6a53d03d1a7da96fe15c076df37f0341fe6
Columns — 15
payee_id
— The payee's key on this site, and the last part of the address of its recipient page. The payee name slugified, then twelve hex characters of an MD5 over the name and full mailing address as FSA printed them, so two payees differ wherever FSA's five fields differ.
payee_name
— The payee as FSA printed it.
mailing_city
— The city of the address FSA mails payments to. Not where any farm is.
mailing_state
— That address's state.
mailing_zip
— That address's ZIP code, as printed.
state_fsa_code
— FSA's code for the state office that made the payment. Not a FIPS code.
state_fsa_name
— That state office's state.
county_fsa_code
— FSA's code for the county office that made the payment. Not a FIPS code.
county_fsa_name
— That county office's county.
program_code
— FSA's accounting program code.
program_description
— FSA's name for the program, as printed.
program_year
— The program year the payment is under, which can differ from the year it was paid.
payment_year
— The calendar year of FSA's file the row is in.
payment_date
— The date of the disbursement.
amount
— The disbursement in dollars. Negative where FSA records a repayment.

Checked before publishing: 22 years' rows and dollars equal the hub's national totals by year; 100 top recipients' keys and yearly totals equal the hub's list of the largest recipients.

Every committed table for Farm program payment files, with its size and a download.
TableRowsSizeDownload
Payments by county and year3,06854 KBCSV
Payments by program and year1,28820 KBCSV
Payments by state and year551 KBCSV
How concentrated the payments are10 KBCSV
Payments nationally, by year221 KBCSV
Payments by state, program and year1,53521 KBCSV
The largest recipients, by total1007 KBCSV
The largest recipients, by number of payments1007 KBCSV
The largest recipients in each state1,37584 KBCSV

What it cannot show

— 8 limitations
  • A year's files are not necessarily that year's farm payments. Measured on the whole 2007 file: 212,296 rows, $74,762,121.50, and three accounting program codes, of which “INCOME LOSS - MILK, PART 2” is 212,275 rows and $74.6 million. One program is not "farm programs" and $74.8 million is not a year of American farm support, so the early files hold one program's payments rather than the whole payment record. The recent years are about thirty times the size per year, which is consistent with a complete record and is not evidence of one; no page may state a year's total until that year's program list has been read.
  • The file carries no recipient identifier. FSA publishes a "Standard Payment File Layout" listing a “Customer Number” — "FSA Assigned Number - Replaces Tax ID" — and that document describes a different file. The published downloads have sixteen columns and no key of any kind. A payee is a formatted name and a mailing address, so two rows are the same payee only if the site decides they are. This site counts rows as one payee where FSA printed the same name and the same full mailing address, so a payee whose address changed is two payees here.
  • One row is one disbursement, not one recipient and not one year. A payee can take several payments under one program on one day, and two rows in the 2004 file do exactly that, differing only in amount. Nothing here deduplicates them.
  • The address is the mailing address FSA holds for payment, not the location of any farm. A payee in Los Banos may farm elsewhere, and an entity may be addressed to an accountant.
  • “Accounting Program Year” and the payment date disagree by design — a 2019 program can pay out in 2020 — so a figure is only a year's figure once the page says which year it means.
  • The publisher has spelled one column two ways. 2008, 2009 and 2010 head the street “Delivery Address”; 2011 onwards heads it “Delivery Address Line”. Sixteen columns either way, same order, same meaning, so it is mapped rather than left blank — the street is one of the five fields a payee is identified by, and a year normalized without it would give the same person a different identity from every other year.
  • State and county are FSA's own codes, not FIPS. 6 is California where FIPS is 06, and the county numbering does not line up either.
  • A payee is not a farm. FSA pays the legal entity on its file, so one operation with three LLCs is three payees and a page must not call any of them a farm.

OPM Federal Workforce Data — Federal Employment Raw Data (EHRI Status)

Publisher
U.S. Office of Personnel Management
License
U.S. Government work, public domain. OPM's own machine-readable catalog at https://www.opm.gov/data.json, read 25 September 2026, gives every one of its 134 datasets "license": "http://www.usa.gov/publicdomain/label/1.0/" — the usa.gov public-domain label — with publisher "U.S. Office of Personnel Management". The five FedScope entries (OPM-OCIO-0027 to OPM-OCIO-0031) carry accessLevel "public". The data is a federal agency's own record of the civilian workforce, a U.S. Government work under 17 U.S.C. 105.
Cadence
Monthly
This copy fetched
Rows held
432,907
Columns
17 columnsas_of_month, measure, parent_code, code, name, employees, employees_pay_withheld, employees_paid, pay_median, pay_mean, pay_p10, pay_p25, pay_p75, pay_p90, pay_min, pay_max, publisher_flagged_incomplete
Pages built on it
Federal salaries
Download
Federal civilian employment and published pay by agency, occupation, state and month — gzipped CSV, 905 KB

What is committed here is a summary. The full table has 432,907 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is f082a84541752d8d14906249208d4cfaab7435fd6c3b4b963d5a0bc3f82eb1c5, so every figure here can be traced to the bytes it came from.

What it cannot show

— 14 limitations
  • Nearly half the workforce has no published salary, and the missing half is not random. In the July 2026 file, 952,126 of 2,020,230 rows — 47.13% — carry the literal string REDACTED where the pay should be. The Department of Defense is 675,945 of 675,945, one hundred percent; Homeland Security 76.0%; Justice 65.9%. Every pay figure here is over employees_paid, which is carried on the same row, and no page may print a pay figure without it.
  • The redaction reaches the map before it reaches the salary. 949,276 rows — 46.99% — have every location field redacted together: state, county, city, country, metropolitan and combined statistical area, and locality pay area. Those employees are kept, under the code REDACTED, so that a geographic measure still sums to the national total and the size of what is missing is a figure on the page rather than an absence.
  • The redaction is OPM's published policy, not a data-quality fault. Its Data Release Policy of December 2024 names ten security and sensitive agencies — the FBI, the Drug Enforcement Administration, the Bureau of Alcohol, Tobacco, Firearms and Explosives, the Alcohol and Tobacco Tax and Trade Bureau, the Secret Service, Customs and Border Protection, Immigration and Customs Enforcement, the Bureau of Prisons, the Homeland Security intelligence office (HSOS04) and every U.S. Attorneys' Office — whose employees outside the Washington area have pay withheld; forwards every Defense request to Defense; and names twenty-six sensitive occupational series, and the file obeys it: series 1811, 1801, 1895, 1896, 0007, 0132, 0083, 0512, 0930, 1169, 0084 and 0082 are each 100% redacted.
  • OPM does not publish every federal employee at all. The intelligence community and some Department of Defense positions are excluded from the file, so a total here is the workforce OPM publishes and not the federal workforce.
  • A month is published about five weeks after it ends: the July 2026 files were published on 2 September 2026, measured from the API's own publish dates. So the newest month available is never the month just gone.
  • There is no name and no employee identifier. All 64 published columns were enumerated on 25 September 2026; the file carries neither. OPM's catalog marks every system that does hold names — the Central Personnel Data File, the EHRI statistical data mart, the electronic Official Personnel Folder, EHRI Status, Dynamics and Payroll — accessLevel "non-public". No page built on this source names a federal employee.
  • Employment is quarterly from 2005-03 to 2011-06 and monthly from 2011-07. All fifty months absent from the range are in the quarterly era. A trend drawn as a monthly line across that break asserts detail the publisher did not supply.
  • Files are revised. Of the 207 current employment months, 24 are at version 1, 51 at version 2 and 132 at version 3, so a month's figure can change between snapshots. The version is part of the file's identity and is recorded in the snapshot name.
  • The publisher flags months whose submissions are incomplete. As of 25 September 2026 it flags June and July 2026: a change in Department of War data processing stopped some components submitting. Measured, Defense falls 684,904 (May 2026) to 675,945 (July), which is a reporting gap and not a workforce change.
  • duty_station_county_code is three digits with no state, so it is not a county identifier on its own. This source assembles the state-county pair once, here, and no query reassembles it.
  • Average salary is an average over a group, never a person's pay, and annualized_adjusted_basic_pay is a rate excluding overtime, awards and bonuses. Pay figures are taken over "PER ANNUM" rows only; the other pay bases are counted separately because an annualized figure for an intermittent appointment is a projection.
  • No pay figure is published for a cell with fewer than five paid rows. A cell of one person published with a median is that person's salary. The cell keeps its headcount and its pay columns are null with the reason stated, never zero.
  • Rows of different measures describe the same employees from different angles. A department row and a state row overlap, and summing across measures double-counts.
  • An agency's name on these pages is OPM's own full name for its code, from the department column of the same file, because the agency column is cut at 40 characters; OPM's label is kept beside it as published. Two names come from elsewhere: the Office of the National Cyber Director, from 6 U.S.C. 1500(a), because OPM's department column drops "National", which its own agency label has; and the Committee for Purchase from People Who Are Blind or Severely Disabled, spelled as OPM's September 2015 file spells it, because the department column has printed "SE VERELY" since September 2017.

U.S. House of Representatives — Statement of Disbursements (payroll)

Publisher
Chief Administrative Officer, U.S. House of Representatives
License
"Pursuant to federal law, government-produced materials appearing on this site are not copyright protected. The United States Government may receive and hold copyrights transferred to it by assignment, bequest, or otherwise." — house.gov, Terms of Use (https://www.house.gov/terms-of-use), read 26 September 2026 and linked from the Statement of Disbursements page itself. The Statement is published by law (the House has been required to publish it since 1964) and is a U.S. Government work under 17 U.S.C. 105.
Cadence
Quarterly
This copy fetched
Rows held
698,030
Columns
16 columnsperiod_start, period_end, organization, organization_code, program, program_code, payee_name, title, perform_start, perform_end, transaction_date, data_source, document, budget_object_code, amount, source_row

What is committed here is a summary. The full table has 698,030 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 289a059e2e77b6babd2745d4fdcf09fde7147dcb97c9e9670fe32b5056539874, so every figure here can be traced to the bytes it came from.

Every committed table for U.S. House of Representatives — Statement of Disbursements (payroll), with its size and a download.
TableRowsSizeDownload
House payroll by printed name and the House’s own program1612 KBCSV
House payroll lines, printed names and offices by quarter421 KBCSV
House payroll by printed name, how each whole year is spread101 KBCSV

What it cannot show

— 7 limitations
  • A name is not an identity. The file carries each employee's name as printed and no identifier: its vendor ID column is blank on every payroll row (April–June 2026, measured). 819 of the 14,187 names in that quarter appear under more than one office. Two people can share a name, and one person can be printed two ways; nothing here decides which, and no page may present two rows as one person without saying how it decided.
  • An amount is what was disbursed in the quarter, not a salary. A row pays for the period printed beside it, which is often the whole quarter and often not: monthly pay, a start or departure mid-quarter, and adjustments reaching back into earlier periods. An annual figure is not the quarter's amount times four, and no page may present it as one.
  • It covers the House of Representatives: Members' offices, committees, leadership and the House's officers. Not the Senate, whose report is a set of signed PDFs, and not the executive branch.
  • The detail grid interleaves detail rows with subtotal and grand-total rows, so summing the file counts every dollar more than once: April–June 2026 sums to $1,829,798,255 naively against a quarter of $457,449,563.64. This source keeps detail rows only, and refuses a file whose detail rows do not equal the publisher's own grand totals to the cent.
  • Only the personnel compensation rows are kept: the payroll. The House's other spending in the same file — travel, supplies, rent, equipment, printing, franked mail — belongs to the spending hub and is not counted here.
  • The files come in two layouts — 12 columns to July–September 2022, 18 from October–December 2022, adding organization, program and object codes — and two encodings: the older files are Windows-1252, not UTF-8. The codes are null before the layout that carries them.
  • An office's name carries the legislative year it belongs to ("2026 HON. JAMES GALLAGHER"), and a quarter's file can carry payments on the previous year's accounts. Both are as published.

IPEDS — every postsecondary institution, with NCES's derived headline measures

Publisher
National Center for Education Statistics, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works. This is our record, not NCES's words: NCES publishes no license for the IPEDS files, and Kian decided on 25 September 2026 to record them as the DCAS source is recorded — NCES is a federal statistical agency and these are its own survey files, U.S. Government works under 17 U.S.C. §105. The evidence, read 25 September 2026, is in this site's register of licenses: the IES public-domain statement names ies.ed.gov only; the IPEDS pages' footer reads "All Rights Reserved", a template line that cannot make a government work copyrightable; the Department's catalog marks IPEDS cc-zero for 2003–04 to 2013–14. NCES's citation is carried on every page built from these files.
Cadence
About three releases a year: the Fall components (directory, enrollment, completions) in July to September, Winter (admissions, graduation, aid) in December, Spring (finance, staff) in January. Provisional first, final a year later at the same address
This copy fetched
Rows held
5,985
Columns
47 columnsunitid, institution_name, city, state, zip, county_fips, county_name, opeid, ein, website, sector, control, level, degree_granting, hbcu, tribal, locale, active, title_iv_flag, in_first_look_universe, carnegie_basic, latitude, longitude, directory_collection, enrollment_12mo, undergraduate_12mo, graduate_12mo, enrollment_collection, enrollment_status, associate_degrees, bachelors_degrees, masters_degrees, completions_collection, completions_status, graduation_rate, bachelors_6yr_rate, graduation_collection, graduation_status, percent_admitted, admissions_yield, admissions_collection, admissions_status, tuition_fees, price_in_state_on_campus, price_out_of_state_on_campus, cost_collection, cost_status
Pages built on it
Schools
Download
Colleges by state and sector, with enrollment and degrees — gzipped CSV, 4 KB

What is committed here is a summary. The full table has 5,985 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is dce58a67f8ce95971c739218c8c3c9529c2320276f37fc654c8163ba7fe319bf, so every figure here can be traced to the bytes it came from.

What it cannot show

— 5 limitations
  • Each measure is for its own year. The directory, 12-month enrollment and completions are from the 2025–26 collection (enrollment and awards during 2024–25); graduation rates, admissions and cost are from 2024–25. A page states each group's collection beside it.
  • A college absent from a derived file has no figure, and says so. Admissions are collected only from colleges that do not admit every applicant, cost only from those that report it, and graduation rates only for cohorts the college has. A null is a status, not a zero.
  • Enrollment is a 12-month unduplicated headcount, everyone enrolled at any point in the year, not fall enrollment and not full-time equivalents.
  • Nonrespondents are imputed by NCES. "Data are imputed for nonresponding institutions" (EFFY2025 dictionary). The derived files do not flag which values are imputed.
  • The universe is NCES's. The First Look universe — Title IV institutions, administrative offices and the service academies, PSET4FLG 1 or 3 — is 5,779 for 2025–26, NCES's own figure. Branch campuses reported by a parent (498 in 2025–26) are not separate rows.

Form 990 series e-file data

Publisher
Internal Revenue Service
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Monthly
This copy fetched
Rows held
760,592
Columns
32 columnsein, return_type, tax_period_end, tax_year, organization_name, total_revenue, total_expenses, contributions, program_service_revenue, investment_income, grants_paid, grants_reported, expenses_program, expenses_management, expenses_fundraising, prior_year_revenue, prior_year_expenses, prior_year_contributions, total_assets_eoy, total_assets_eoy_fmv, is_amended, officer_count, listed_people, highest_paid_name, highest_paid_title, highest_paid_role, highest_paid_compensation, highest_paid_reportable_from_org, highest_paid_reportable_from_related, highest_paid_other_compensation, highest_paid_is_person, object_id
Pages built on it
Nonprofits

What is committed here is a summary. The full table has 760,592 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is ff1be60b86f6e53ed0f0d302a63fa1196922cc90cc6afd1712c07b5e7c873c10, so every figure here can be traced to the bytes it came from.

Every committed table for Form 990 series e-file data, with its size and a download.
TableRowsSizeDownload
Filers by category272 KBCSV
Filers by the form they filed30 KBCSV
Filers by state614 KBCSV
Filers by tax-exempt subsection312 KBCSV
Private foundations by state584 KBCSV
Private foundations nationally10 KBCSV
Filers nationally10 KBCSV
The largest filers in each category67591 KBCSV
The largest filers in each state1,435180 KBCSV
The largest private foundations10011 KBCSV
The largest grantmakers10015 KBCSV
The highest reported compensation10012 KBCSV
The largest filers by assets10015 KBCSV
The largest filers by revenue10015 KBCSV

This is a subset. The most recent filing on record for each organization, from the three most recent years of batches. Earlier filings are not kept, and the committed copy is the aggregates and ranked lists the published pages read rather than the 752,096 organizations themselves.

What it cannot show

— 6 limitations
  • A Form 990 is self-reported by the organization and is not audited by the IRS before publication. Figures are what the organization said, not a verified account.
  • Filings lag the year they describe, usually by one to two years. The most recent filing on record is rarely the most recent year of activity.
  • Only electronically filed returns appear here. Paper filings are not in this dataset.
  • "Revenue" on a Form 990 includes contributions, grants, program service revenue, investment income and more. It is not comparable to a company's sales.
  • Reported compensation covers the organization's fiscal year and may include amounts from related organizations. It is what the organization reported paying, not a verified salary, and it is not what a person took home.
  • Organizations restate and amend returns. A figure is only as current as the snapshot it was drawn from.

Exempt Organizations Business Master File

Publisher
Internal Revenue Service
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Monthly
This copy fetched
Rows held
1,975,975
Columns
17 columnsein, name, city, state, subsection, ruling_year, ntee_code, asset_code, income_code, filing_requirement, foundation_code, pf_filing_requirement, deductibility_code, organization_form_code, asset_amount, income_amount, revenue_amount

What is committed here is a summary. The full table has 1,975,975 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is d739b1fed31168fc6223611452cdaad3b00e9c76c9ebb760eec327b609336c26, so every figure here can be traced to the bytes it came from.

Every committed table for Exempt Organizations Business Master File, with its size and a download.
TableRowsSizeDownload
Registered organizations by category260 KBCSV
Registered organizations by state621 KBCSV
Registered organizations by state and category1,4646 KBCSV
Registered organizations by tax-exempt subsection340 KBCSV
Registered organizations nationally10 KBCSV
Registered organizations with a Form 990 on record167,9316.0 MBCSV

This is a subset. The organizations the IRS places in an income bracket of $1,000,000 or more — 167,682 of 1,967,361. The whole file is 54 MB compressed and is republished monthly, so it lives in object storage and the database rather than in the repository.

What it cannot show

— 9 limitations
  • The BMF records registration status, not activity. An organization appears here because the IRS recognizes its exemption, not because it is operating.
  • Nothing in this file is revoked or terminated. The only exempt-organization status codes present are 01, 02, 12 and 25, all of which mean the exemption stands; an organization whose exemption is revoked leaves the file entirely and appears on the IRS's separate Automatic Revocation of Exemption List. So "currently exempt" here is a statement about the snapshot date, and an organization vanishing between refreshes is a finding rather than a fault.
  • The income and asset codes are brackets, and bracket 0 is not "nothing". It means the IRS has no Form 990 series return to draw a figure from, which is true of about 64% of the file — mostly Form 990-N postcard filers and organizations not required to file at all. Treating bracket 0 as $0 would report an income of zero for most organizations in the country.
  • Churches and other organizations that are not required to apply for recognition, and that have not applied, are absent. The IRS states this in its own dictionary for this extract, alongside self-declared organizations and split-interest trusts. The church count here is the churches that chose to apply, and it is not a count of churches.
  • Nine organizations carry subsection 82, which is IRC 527 — political organizations, not exempt under 501 at all. They are campaign committees. Any page describing this file as tax-exempt organizations has to exclude them or name them for what they are.
  • Asset and income are bracket codes, not reported amounts, for most rows. Where dollar columns are populated they come from the most recent return the IRS has processed, which may be one or two years old.
  • The NTEE code is assigned at registration and is often not updated when an organization's work changes.
  • Churches and certain affiliated organizations are exempt without applying, so many do not appear in this file at all.
  • The IRS republishes the file monthly and organizations are added and removed. A figure is only as current as the snapshot it was drawn from.

State of Minnesota — State Employee Payroll Data

Publisher
Minnesota Management and Budget
License
The payroll page (https://mn.gov/mmb/transparency-mn/payrolldata.jsp), as Kian saved it from a browser on 3 October 2026 (payrolldata-page-2026-10-03.htm), because mn.gov answers this project's requests with a bot challenge (Radware, 2 October 2026), which was not worked around. It states no license, terms or restriction; it describes the files as public: "There are a total of 48 data elements about state employees that are public. The downloadable data set contains all 48 data elements." And: "Finally, the complete data files with information for all state employees can be downloaded for additional research and queries." The mn.gov portal's policy page (https://mn.gov/portal/about-minnesota/about-mn-gov.jsp), read 2 October 2026 through the shared fetch client, under "Use of Information on this Site": "The state agency pages are presented as an informational service and any content should be independently verified with specific agencies when an official record of action is required. State agency authored documents are in the public domain. Copyright and access restrictions apply." And under "Disclaimer": "Neither the State of Minnesota, nor any agency, officer, or employee of the State of Minnesota warrants the accuracy, reliability or timeliness of any information published by this system, nor endorses any content, viewpoints, products, or services linked from this system, and shall not be held liable for any losses caused by reliance on the accuracy, reliability or timeliness of such information. Portions of such information may be incorrect or not current. Any person or entity that relies on any information obtained from this system does so at his or her own risk." The workbooks are MMB's own, a state agency authored document.
Cadence
Annually
This copy fetched
Rows held
2,215,138
Columns
43 columnsfiscal_year, fiscal_year_end, record_type, temporary_id, record_nbr, listing, employee_name, agency_nbr, agency_name, department_nbr, department_name, branch_code, branch_name, job_code, job_title, location_nbr, location_name, location_postal_code, location_county_name, reg_temp_code, reg_temp_desc, classified_code, classified_desc, original_hire_date, last_hire_date, job_entry_date, full_part_time_code, full_part_time_desc, salary_plan_grid, salary_grade_range, max_salary_step, compensation_rate, comp_frequency_code, comp_frequency_desc, position_fte, bargaining_unit_nbr, bargaining_unit_name, active_on_june_30, regular_wages, overtime_wages, differentials_wages, other_wages, total_wages

What is committed here is a summary. The full table has 2,215,138 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is ee50cf0bdd9783c15c67d06013c5fc96425cd9c0ca4d8ae0dece37f4928f4ab3, so every figure here can be traced to the bytes it came from.

Every committed table for State of Minnesota — State Employee Payroll Data, with its size and a download.
TableRowsSizeDownload
People and median pay by agency in each fiscal year1,50852 KBCSV
MMB numbers in each numbering, and how many have a page offered to search engines40 KBCSV
The hundred people paid the most in the newest fiscal year, named1005 KBCSV
People paid, pay and its spread in each fiscal year, and how the year’s pay is read161 KBCSV

What it cannot show

— 8 limitations
  • Earnings are a fiscal year's expenditure on a person, not a salary. MMB: "the dollars included in the salary column represents the total amount expended on behalf of the individual for regular work, vacation, sick, holiday etc. … These amounts may also include retroactive earnings for the prior fiscal year. This amount does not represent what would appear on their W-2 … The figure does not necessarily reflect the employee's current or annualized salary." A part-year employee's earnings are part of a year.
  • Earnings and positions are two records, and MMB says they do not always match. "While the Earnings data is summed up by individual, the HR data represents all currently active positions the individual held as of June 30 … An individual may have earned their previous year's earnings in a completely different position or set of positions sometimes crossing multiple agencies." A person's earnings are not attributed here to any one position or agency.
  • temporary_id identifies a person within a run of years, not across the series. MMB: "a unique number created just for this report and is not an extract from the database." Measured over all sixteen workbooks: within FY2011–2015, FY2016–2017, FY2018–2019 and FY2020–2026 a number keeps its printed name from year to year (for example 54,978 of 54,987 names printed once in both FY2016 and FY2017), and at FY2016, FY2018 and FY2020 every number was reissued, with none in both years. So one person can be several numbers over the series, and nothing here links them.
  • The wage columns changed twice. FY2011 publishes salary, overtime, differentials and other pay in four columns, and its other pay includes "relocation expenses, clothing/uniform allowance, communication allowance"; FY2012–2015 publish regular, overtime and other wages, with differentials inside other wages; FY2016 on add a total-wages column and say allowances "are no longer included". The table carries a total only where MMB prints one. The pages read a year's pay as that total, and for FY2011–2015 as the printed columns added, which is how MMB defines its total; every FY2016–2026 total equals its columns to the cent.
  • Four numbers are two people. Three in FY2018 and one in FY2019 are listed twice in the Earnings sheet with different wages, and each number's two positions print two different names in two agencies: MMB gave two people one number. Both Earnings rows are kept as published, numbered by listing; which wages are whose, the file does not say, so neither is attributed to either name.
  • FY2021's Earnings sheet prints "-" in wage cells with nothing in them (41,794 of 63,016 overtime cells); a dash is read as no value, not as zero.
  • Not every state employee. MMB: "The database includes all state employees in the Executive and Judicial branches, constitutional offices including the Attorney General, State Auditor, and Secretary of State's. The data also includes the salary data for employees of colleges and universities within Minnesota State … However, the information does not include the University of Minnesota, the Metropolitan Council, Historical Society, or other semi-state agencies funded by grants that are not on the state payroll system. Certain small groups of individuals are also excluded - an example is special agents employed by the Department of Public Safety and a limited number of other identity-protected law enforcement individuals."
  • FY2011's workbook overlaps the calendar-year 2010 file it replaced: "The overlap is for the period July 1, 2010 - December 31, 2010." Only the fiscal-year workbooks are read here.

NCES Common Core of Data — public school districts, newest full school year

Publisher
National Center for Education Statistics, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works. This is our record, not NCES's words: NCES publishes no license for the CCD files, and Kian decided on 25 September 2026 to record them as the DCAS source is recorded — NCES is a federal statistical agency and these are its own survey files, U.S. Government works under 17 U.S.C. §105. The evidence is in this site's register of licenses. NCES's requested citation is carried on every page built from these files: "SOURCE: U.S. Department of Education, National Center for Education Statistics, Common Core of Data (CCD)", with the file and version.
Cadence
Annual. A preliminary directory in the summer, the school year's other files about fifteen months after it begins (NCES's release calendar)
This copy fetched
Rows held
19,630
Columns
25 columnsleaid, district_name, state, state_fips, school_year, status_code, status, operational, district_type_code, district_type, charter_lea, supervisory_union, nyc_supervisory_component, grade_low, grade_high, city, zip, website, operational_schools, enrollment, enrollment_status, staff_fte, staff_fte_status, teachers_fte, teachers_fte_status
Pages built on it
Schools

What is committed here is a summary. The full table has 19,630 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is a0390d85db4ceff58fa4b25c29e374289dbd5c4062d38e0e18ee39849e8b853b, so every figure here can be traced to the bytes it came from.

Every committed table for NCES Common Core of Data — public school districts, newest full school year, with its size and a download.
TableRowsSizeDownload
Every operational public school district, with its pupils and teachers19,400347 KBCSV
Operational public school districts by state, as listed and as NCES counts them571 KBCSV

What it cannot show

— 4 limitations
  • NCES counts New York City as one district. Its 33 geographic districts are components of a supervisory union, and NCES's own count of districts (19,183 in the fifty states and DC for 2024–25) counts them once. They are rows here, marked, so a count can be taken either way and say which.
  • Not every district enrolls students. Service agencies, supervisory unions and state agencies are districts in CCD; 18,548 of 19,630 carry a membership total in 2024–25.
  • A count NCES did not publish keeps NCES's reason, missing, suppressed or not reported, and a district absent from a file is marked so. A zero is a zero.
  • Teachers are NCES's derived major staffing category where it derived one, and the state's reported figure otherwise; the status says which.

NCES School District Finance Survey (F-33) — newest fiscal year

Publisher
National Center for Education Statistics, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works. This is our record, not NCES's words: NCES publishes no license for the CCD files, and Kian decided on 25 September 2026 to record them as the DCAS source is recorded — U.S. Government works under 17 U.S.C. §105. The F-33 documentation itself says "Unless specifically noted, all information contained herein is in the public domain" (NCES 2025-306). The evidence is in this site's register of licenses. NCES's requested citation is carried on every page built from this file.
Cadence
Annual, about two years after the fiscal year: a provisional file (version 1a), and a final one (2a) a year later, at the same time as the next year's provisional
This copy fetched
Rows held
19,570
Columns
21 columnsleaid, district_name, state, fiscal_year, release, membership, membership_status, total_revenue, total_revenue_status, federal_revenue, federal_revenue_status, state_revenue, state_revenue_status, local_revenue, local_revenue_status, total_expenditure, total_expenditure_status, current_expenditure, current_expenditure_status, capital_outlay, capital_outlay_status
Pages built on it
Schools
Download
Every school district’s revenue by source, spending and pupils, one fiscal year — gzipped CSV, 628 KB

What is committed here is a summary. The full table has 19,570 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 2a4bddf8a04aff670babc8f16bc480020cb0c8d3de1b4964e06921bf7822b61d, so every figure here can be traced to the bytes it came from.

What it cannot show

— 4 limitations
  • Provisional until final. The newest year is published as version 1a, provisional, and revised as 2a a year later; the release column says which a figure is from.
  • Two years behind. FY 2023, the 2022–23 school year, is the newest NCES has published in September 2026.
  • A negative number in the file is a code, not a figure. Missing (-1), not applicable (-2), suppressed membership (-3) and below NCES's quality standard (-9) are nulls here, with the reason kept.
  • Per-pupil spending needs a denominator a page names. Current expenditure divided by fall membership is NCES's per-pupil measure; districts with no membership (service agencies) have none, and a page states which districts a median is over.

NCES School District Finance Survey (F-33) — every fiscal year in the current format

Publisher
National Center for Education Statistics, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works. This is our record, not NCES's words: NCES publishes no license for the CCD files, and Kian decided on 25 September 2026 to record them as the DCAS source is recorded — U.S. Government works under 17 U.S.C. §105. The F-33 documentation itself says "Unless specifically noted, all information contained herein is in the public domain" (NCES 2025-306). The evidence is in this site's register of licenses. NCES's requested citation is carried on every page built from this file.
Cadence
Annual, about two years after the fiscal year: a provisional file (version 1a), and for some years a final one (2a) a year later
This copy fetched
Rows held
367,012
Columns
22 columnsleaid, district_name, state, fiscal_year, version, release, membership, membership_status, total_revenue, total_revenue_status, federal_revenue, federal_revenue_status, state_revenue, state_revenue_status, local_revenue, local_revenue_status, total_expenditure, total_expenditure_status, current_expenditure, current_expenditure_status, capital_outlay, capital_outlay_status
Pages built on it
Schools
Download
School district revenue, spending and pupils by state, every fiscal year from 2005 — gzipped CSV, 45 KB

What is committed here is a summary. The full table has 367,012 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 3baf41c53c94182374914a497b236fa33e2010b7dbb232d45c022570789d2e83, so every figure here can be traced to the bytes it came from.

What it cannot show

— 5 limitations
  • Nineteen years, not decades. The F-33 goes back to FY 1990, but only from FY 2005 does the file say which values are missing; before that a gap is written as zero. The series starts where a zero can be believed.
  • Dollars as reported, not adjusted for inflation. Every figure is in the year's own dollars. A trend that compares FY 2005 with FY 2023 says so, or adjusts with a named price index.
  • The latest version NCES serves, which is not always final. FY 2005–FY 2018 and FY 2023 are provisional files; FY 2019–FY 2022 are final. A final file differs from the provisional one NCES's First Look was written from: FY 2021's per-pupil figure is $14,147 in NCES 2024-303 and $14,185 here.
  • A negative number in the file is a code, not a figure. Missing (-1), not applicable (-2), suppressed membership (-3) and below NCES's quality standard (-9) are nulls here, with the reason kept.
  • The set of districts changes from year to year. Districts open, merge and close, and agencies without a nonfiscal LEAID carry a dummy one from 2007 (a D in the third place). A state's sum is over the districts in that year's file; a district's own series can have gaps.

NCES Common Core of Data — public schools, newest full school year

Publisher
National Center for Education Statistics, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works. This is our record, not NCES's words: NCES publishes no license for the CCD files, and Kian decided on 25 September 2026 to record them as the DCAS source is recorded — NCES is a federal statistical agency and these are its own survey files, U.S. Government works under 17 U.S.C. §105. The evidence, read 25 September 2026, is in this site's register of licenses: NCES's copyright URL leads to an IES public-domain statement that names ies.ed.gov only; the IPEDS pages' footer reads "All Rights Reserved"; the Department's catalog marks CCD cc-zero for 1986–87 to 2015–16 only. NCES's requested citation is carried on every page built from these files: "SOURCE: U.S. Department of Education, National Center for Education Statistics, Common Core of Data (CCD)", with the file and version.
Cadence
Annual. A preliminary directory in the summer, the school year's other files about fifteen months after it begins (NCES's release calendar)
This copy fetched
Rows held
102,178
Columns
33 columnsncessch, leaid, school_name, district_name, state, state_fips, school_year, status_code, status, operational, school_type, charter, level, grade_low, grade_high, city, zip, website, nslp_status, virtual_instruction, county_fips, county_name, locale, latitude, longitude, enrollment, enrollment_status, teachers_fte, teachers_fte_status, free_reduced_lunch, free_reduced_lunch_status, direct_certification, direct_certification_status
Pages built on it
Schools

What is committed here is a summary. The full table has 102,178 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is f8769dd5dcda6251480617b73de40f3b9b295f7d2fefab44dbbee8c79dd580cc, so every figure here can be traced to the bytes it came from.

Every committed table for NCES Common Core of Data — public schools, newest full school year, with its size and a download.
TableRowsSizeDownload
Operational public schools and pupils by state and locale2192 KBCSV
Operational public schools, pupils, teachers and lunch counts by state573 KBCSV

What it cannot show

— 5 limitations
  • A school with no count in a file has a reason, and it is shown. NCES writes a missing, not-applicable or suppressed count as a null with a flag, and a school absent from a file is absent. Tennessee reported no teacher FTE for 2024–25; Pennsylvania's and South Carolina's free and reduced-price lunch counts are suppressed statewide.
  • Schools' totals do not add to NCES's state totals. In 2024–25 the school enrollment counts sum to 49,034,047 for the fifty states and DC, and NCES's state file says 49,269,882. 817 operational schools have no membership row. A figure summed from schools says what it is over.
  • Enrollment is a count on one day in the fall, as each state reports it, not attendance and not the year's total. California and Oregon's pre-kindergarten counts are, in NCES's words, "no longer imputed, starting with SY 2021–22".
  • Free and reduced-price lunch eligibility is not a count of poor children. Schools under the Community Eligibility Provision serve every student free and may report every student, and direct certification counts students matched to benefit programs. NCES publishes both; neither is an income measure.
  • A school's county is EDGE's, the county its address geocodes to, which is not always the county its district serves.

NCES Common Core of Data — public school membership by state, 2014–15 onwards

Publisher
National Center for Education Statistics, U.S. Department of Education
License
U.S. Government work, license https://www.usa.gov/government-works. This is our record, not NCES's words: NCES publishes no license for the CCD files, and Kian decided on 25 September 2026 to record them as the DCAS source is recorded — NCES is a federal statistical agency and these are its own survey files, U.S. Government works under 17 U.S.C. §105. The evidence is in this site's register of licenses. NCES's requested citation is carried on every page built from these files: "SOURCE: U.S. Department of Education, National Center for Education Statistics, Common Core of Data (CCD)", with the file and version.
Cadence
Annual. A school year's state file about ten months after it begins (NCES's release calendar)
This copy fetched
Rows held
11,080
Columns
7 columnsstate, state_fips, state_name, school_year, grade, membership, membership_status
Pages built on it
Schools
Download
Public school membership by state and school year, with and without pre-kindergarten — gzipped CSV, 8 KB

What is committed here is a summary. The full table has 11,080 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 2021436ca82d3d417ba072292a01ce5e38ddd85d965df89569af6dd89093c47a, so every figure here can be traced to the bytes it came from.

What it cannot show

— 5 limitations
  • From 2014–15 only. Before it, NCES's state file is a different survey file (st131a and older) with its own layout, and this source does not read it. 2014–15 and 2015–16 are the component file in its wide layout, and every year since in its long one; both are read to the same rows.
  • The published national figure can be higher than the states' sum. The Digest of Education Statistics (2024, Table 203.20) matches these totals for every state and year from fall 2014 to fall 2023 except California (fall 2015 and 2019–2023) and Oregon (fall 2015, 2016 and 2020–2023), where it "Includes imputations for nonreported prekindergarten enrollment". This file carries what the states reported, and those years' pre-K is not reported.
  • Two totals, and the pages use the one without adult education, which is the figure NCES publishes as a state's enrollment. The total with it is kept; the difference is a few thousand adults in a handful of states.
  • A count NCES did not publish keeps NCES's reason: missing, not applicable or suppressed (–1, –2 and –9 in the wide layout, NCES 2016-077), and NCES's own flag in the long one. A zero is a zero.
  • Membership is the October 1 count of pupils on a school's rolls, including pre-K where a state reports it; it is not attendance, and it is not the district or school files' sums, which differ slightly from the state's own total.

NCUA credit-union sites — Credit Union Branch Information (the newest quarter)

Publisher
National Credit Union Administration
License
U.S. Government work. This is our record, not the NCUA's words: Kian decided on September 26, 2026 to record the NCUA's data as NCES and IPEDS are recorded — the NCUA is an independent agency of the United States, and the Call Report files are its own compilation of the reports credit unions file with it, U.S. Government works under 17 U.S.C. 105. The publisher's own label, from its catalog https://ncua.gov/data.json ("National Credit Union Administration Enterprise Data Inventory", DCAT-US 3.0, modified September 17, 2026, read September 26, 2026), on the entry "Credit Union and Call Report Data" ("Data collected through 5300 Natural Person Call Report processes."): "license": "https://www.usa.gov/government-works", "accessRights": "Public", the rights statement "This data asset is a public data asset eligible for disclosure in the Federal Data Catalog. It is not exempt from disclosure under FOIA. It is subject to open format and open license requirements.", and the use restriction "Unrestricted", "Use of this dataset is not restricted." The quarterly-data page (last modified 09/14/26): "All data listed on this page is final data." The NCUA's website policies (https://ncua.gov/about/open-government/website-policies, last modified 04/15/26) state no copyright position. The conditions, in this site's register of licenses: no NCUA seal, nothing "official", the NCUA named as publisher with the quarterly file linked.
Cadence
Quarterly. The branch file is in each quarter's Call Report zip, which the NCUA posts about ten weeks after the quarter ends (June 30, 2026's is dated September 14, 2026).
This copy fetched
Rows held
22,837
Columns
21 columnscu_number, site_id, cycle_date, join_number, cu_name, site_name, site_type, main_office, address, address_line2, city, state_abbr, zip, county_name, country, phone, hours, member_services, atm, drive_through, shared_service_center_network
Pages built on it
Banks and credit unions
Download
Credit-union sites by state, the newest quarter — gzipped CSV, 1 KB

What is committed here is a summary. The full table has 22,837 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 2b80532f9daaefc02a62757109fada1830dd0486b0258c0a384ebcaf289e9d45, so every figure here can be traced to the bytes it came from.

What it cannot show

— 5 limitations
  • Sites, not deposits. The NCUA collects shares, deposits and members for a credit union as a whole, not for its offices, so no figure here can be summed by county or city the way the FDIC's Summary of Deposits can.
  • The newest quarter only. This is the list of sites the NCUA had at the quarter's end. A site that closed is simply absent; the file says nothing about when a site opened or closed.
  • The county is a name, not a code, as the NCUA writes it, blank on 52 sites (50 of them abroad). A county page joins it to a county by name within the state and says what did not resolve.
  • Fifty sites are recorded abroad (Germany 16, Japan 13, Italy 5 and others), with no state: most are military credit unions' offices on bases, and at least one is not abroad at all (Tippecanoe's second site is its Lafayette, Indiana, address with the country written as Afghanistan). They are listed on their credit union's page as the NCUA records them and counted in no state.
  • A site's type and flags are the credit union's own report. 5,344 sites are "Corporate Office"; 2,152 are marked as offering no member services. Opening hours are free text, blank on 619 sites.

NCUA 5300 Call Report — quarterly data (every credit union, every quarter from March 1994)

Publisher
National Credit Union Administration
License
U.S. Government work. This is our record, not the NCUA's words: Kian decided on September 26, 2026 to record the NCUA's data as NCES and IPEDS are recorded — the NCUA is an independent agency of the United States, and the Call Report files are its own compilation of the reports credit unions file with it, U.S. Government works under 17 U.S.C. 105. The publisher's own label, from its catalog https://ncua.gov/data.json ("National Credit Union Administration Enterprise Data Inventory", DCAT-US 3.0, modified September 17, 2026, read September 26, 2026), on the entry "Credit Union and Call Report Data" ("Data collected through 5300 Natural Person Call Report processes."): "license": "https://www.usa.gov/government-works", "accessRights": "Public", the rights statement "This data asset is a public data asset eligible for disclosure in the Federal Data Catalog. It is not exempt from disclosure under FOIA. It is subject to open format and open license requirements.", and the use restriction "Unrestricted", "Use of this dataset is not restricted." The quarterly-data page (last modified 09/14/26): "All data listed on this page is final data." The NCUA's website policies (https://ncua.gov/about/open-government/website-policies, last modified 04/15/26) state no copyright position. The conditions, in this site's register of licenses: no NCUA seal, nothing "official", the NCUA named as publisher with the quarterly file linked; and the names of people in the NCUA's files are never read.
Cadence
Quarterly. The NCUA posts a quarter's file about ten weeks after it ends (June 30, 2026's zip is dated September 14, 2026) and its Quarterly Credit Union Data Summary about ten days later (24 September). This source loads a quarter once both are published, because the summary is what the load is held to.
This copy fetched
Rows held
860,487
Columns
23 columnscu_number, cycle_date, join_number, partial_cycle, name, city, state_abbr, zip, county_fips, charter_state, cu_type, federally_insured, rssd, year_opened, minority_depository, total_assets, shares_and_deposits, loans_and_leases, members, net_worth, cecl_transition_provision, net_income_ytd, branches
Pages built on it
Banks and credit unions

What is committed here is a summary. The full table has 860,487 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 8c9832414c71daca6feb191a799a5ec2f2ae77eb213295c43d7edda29238c2e9, so every figure here can be traced to the bytes it came from.

Every committed table for NCUA 5300 Call Report — quarterly data (every credit union, every quarter from March 1994), with its size and a download.
TableRowsSizeDownload
Credit unions, members, assets and shares by quarter since 1994, insured or not2608 KBCSV
Credit unions, members, assets and shares by state, the newest quarter632 KBCSV
The 100 largest credit unions by assets and by members, each ranked among all that filed1145 KBCSV

What it cannot show

— 6 limitations
  • Before September 2002, March and September hold only the credit unions that filed that quarter. The March 1994 file has 1,144 credit unions against 12,363 in June 1994. Those 17 quarters are flagged partial_cycle; a series across quarters uses June and December until June 2002 and every quarter from September 2002. The files do not say which credit unions filed quarterly then.
  • Net income is for the year to date. March's figure is one quarter's income and December's is the year's; a quarter's own income is the difference from the previous quarter of the same year. Net income is not reported before September 1999, net worth before December 2000 and branches before March 2003.
  • These are the reports credit unions filed, not audited statements, and credit unions amend them. The files on the NCUA's page carry the amendments; the NCUA's own summary of an earlier quarter was computed before them, so the two can differ by a few tenths of a percent (0.26% on net worth at December 2009, the largest measured apart from December 2008, which the NCUA revised).
  • The address is the mailing address. City and state are where the credit union receives mail (a post-office box on some rows), and its county is the NCUA's code for it, which in two of June 2026's rows is in a different state from the mailing address. A credit union's branches are in the NCUA's branch file, not here.
  • 85 credit unions are not federally insured, and the NCUA's own system figures leave them out. A total over every row is the 4,299; the NCUA's is the 4,214 federally insured.
  • The file does not say why a credit union stops appearing. One that merged or was liquidated is simply absent from the next quarter. One that converted its charter (federal to state, or back) appears under a new charter number; join_number is the same on both, and is what links them.

State of New Jersey — YourMoney Agency Payroll

Publisher
New Jersey Department of the Treasury, Office of Management and Budget
License
The dataset states no license. data.nj.gov's footer links the State of New Jersey's Legal Statement (https://www.nj.gov/nj/legal.shtml), read 26 September 2026, whose Section F, "Copyright and Trademark Limitations", reads: "The State of New Jersey has made the content of these pages available to the public and anyone may view, copy or distribute State information found here without obligation to the State, unless otherwise stated on particular material or information to which a restriction on free use may apply." Nothing on this dataset states such a restriction. Section G restricts the State seal, which is never used.
Cadence
Quarterly
This copy fetched
Rows held
1,380,036
Columns
19 columnscalendar_year, calendar_quarter, as_of_date, payroll_id, last_name, first_name, middle_initial, full_name, original_employment_date, department, section, title, employee_relations_group, compensation_method, rate, ytd_regular_pay, ytd_overtime_pay, ytd_other_pay, ytd_earnings
Pages built on it
State payrolls

What is committed here is a summary. The full table has 1,380,036 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 88635ad27c4af40bb3b9b9bc5c4f9f3a6e6f1fcde9f73849fdb29e82e781d74f, so every figure here can be traced to the bytes it came from.

Every committed table for State of New Jersey — YourMoney Agency Payroll, with its size and a download.
TableRowsSizeDownload
Employees and pay by department in each release64633 KBCSV
Payroll IDs, and how many are in more than one release or were paid $100,000 or more in one10 KBCSV
Employees, pay and its spread in each release, the fourth quarter of every year from 2010 and the newest quarter171 KBCSV
The hundred employees paid the most in the newest whole year, named1004 KBCSV
The hundred employees paid the most overtime in the newest whole year, named1004 KBCSV

What it cannot show

— 5 limitations
  • Earnings are year-to-date, not annual. Every earnings column is the calendar year to the release's date, so an April–June release is half a year, and the fourth-quarter releases are the only complete years. rate is the pay rate on its own basis — annual salary, hourly, per diem or part time — which compensation_method names; 1,525 of the 75,135 employees in April–June 2026 have none listed.
  • It covers New Jersey's state agencies, not its authorities, which the State publishes as a separate dataset, and not its public universities, counties, municipalities or school districts. A state employee's pay here is what the State paid them, and nothing else.
  • The publisher releases the year's fourth quarter for every year from 2010 and only the newest quarter of the current year, so there is no quarterly history before it: a trend is a point a year.
  • payroll_id is the publisher's own key and it is stable across releases, measured rather than assumed: of 75,135 ids in April–June 2026, 71,589 are in October–December 2025 and 71,335 of those carry the same name. The publisher describes it as randomly generated; it identifies a record, and the name printed beside it is the person.
  • The State publishes the literal "none" where an employee has no first name, and "null" where a date is missing. Both are read as null here.

State of Oklahoma Payroll

Publisher
Oklahoma Office of Management and Enterprise Services
License
Each dataset page's "License" field, read 2 October 2026: calendar 2010 to FY2024, fifteen datasets, "License Creative Commons Attribution", linking http://www.opendefinition.org/licenses/cc-by; FY2025, FY2026 and FY2027 "License No License Provided". The portal's About page (https://data.ok.gov/about), read the same day: "By posting state government data in raw, machine-readable formats, it can be reformatted, repurposed, and reused in different ways, allowing the community at large to build custom applications in order to analyze and display the information." oklahoma.gov's Copyright and Protection Acts page is a DMCA procedure, with no reuse restriction. The CC BY datasets are used under CC BY, crediting OMES; the three silent ones as public records under the Oklahoma Open Records Act (51 O.S. § 24A.1 and after), with the same credit (Kian, 2 October 2026). Every dataset page is stored with its files, so the field as it read at each fetch is kept.
Cadence
Monthly
This copy fetched
Rows held
29,579,857
Columns
16 columnssource_dataset, source_file, line, report_date, agency_number, agency_name, last_name, first_initial, middle_initial, hours, amount, check_date, account, account_description, update_date, row_id

What is committed here is a summary. The full table has 29,579,857 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 0582071fa576c65f9e24cf8c045c498a933a1fdf02c89d34cec7e5dfcfb02f5c, so every figure here can be traced to the bytes it came from.

Every committed table for State of Oklahoma Payroll, with its size and a download.
TableRowsSizeDownload
Pay lines and what they paid by pay type in each month3,77063 KBCSV
Pay lines and what they paid by agency in each month28,787480 KBCSV
Pay lines and what they paid in each month1944 KBCSV

What it cannot show

— 5 limitations
  • A name is a surname and two initials, and it is not a person. No file carries an employee number or a first name (2015's column is headed "First Name" and holds an initial), so everyone who shares a surname and initials shares a name here. Nothing joins two lines as one person.
  • Each month is read from the one file named for it. Calendar 2020's dataset carries fiscal Q1 2021, FY2021's carries files calendar 2020 also lists, and FY2026 and FY2027 both carry July 2026; a file listed twice is read once. From FY2022 each monthly file also carries earlier months, and those rows are dropped: a month's rows are its own file's. The file named "calendar-year-2015" holds January to May 2015 only, the q1-2015 file's rows; it is mislabelled and not read, so calendar 2015 is its four quarters. A month no file is named for is read from the one file in its period that carries it: April and May 2020 from OMES's single corrected file for April to June 2020.
  • FY2022 is not shown. OMES's monthly files for that year restate earlier months and do not agree with each other; we have asked OMES. It is loaded, under the same rule, for measurement only, and no page, trend or year-over-year figure reads it.
  • OMES: "The payroll data represents the amount paid to an employee during the reported time period. In addition to regular pay, these amounts may include other pay types such as overtime, longevity, shift differential or terminal pay. This amount does not include any state share costs associated with the payroll i.e. FICA, state share retirement, etc. This amount may vary from an employee's 'salary' due to pay adjustments or pay period timing."
  • Agency numbers are printed with and without a leading zero ("1000" and "01000") in different files, and are carried as printed.

Open Payments — payments grouped by physician

Publisher
Centers for Medicare & Medicaid Services
License
U.S. Government work. CMS declares https://www.usa.gov/government-works as the license on every one of its 74 Open Payments datasets, with accessLevel: public. That page states a government work is "something created by a U.S. government officer or employee as part of their official duties" and is generally not protected by copyright, subject to conditions that do bind us: "You cannot use government materials in a way that implies endorsement by a government agency, official, or employee", and "You also cannot use federal government trademarks or federal government agency logos without permission."
Cadence
Annual
This copy fetched
Rows held
7,458,173
Columns
28 columnsprofile_id, program_year, is_all_years, recipient_type, npi, first_name, middle_name, last_name, name_suffix, city, state, zip_code, country, specialty, license_state, general_total_payment, research_total_payment, invested_total_amount, interest_total_amount, general_total_transactions, research_total_transactions, invested_total_transactions, interest_total_transactions, associated_research_payments, associated_research_transactions, disputed_transactions, undisputed_transactions, has_multiple_ids
Pages built on it
Health

What is committed here is a summary. The full table has 7,458,173 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is a6ddeecffec7dea55aebd4b7b5131ac4a5c5e92f97457d63aa39c1e3a16cb49c, so every figure here can be traced to the bytes it came from.

Every committed table for Open Payments — payments grouped by physician, with its size and a download.
TableRowsSizeDownload
Payments by specialty group70 KBCSV
Payments by state591 KBCSV
Payments by state and program year4126 KBCSV
Payments by program year70 KBCSV
Payments nationally10 KBCSV
The clinicians paid most in each state58923 KBCSV
The clinicians paid most in general payments1005 KBCSV
The clinicians with the largest ownership interests1005 KBCSV
The clinicians with the most research funding1005 KBCSV

What it cannot show

— 6 limitations
  • The file carries a summary row per physician and it looks exactly like a year. Program_Year takes the literal value "All", and that row is the sum of the physician's year rows — checked on the fixture: 20.00 + 93.47 + 104.15 = 217.62, and the All row reads 217.62. So a total computed across this table without filtering is exactly twice the truth, a ranking built on it ranks correctly for the wrong reason, and neither error looks wrong. is_all_years exists so no query has to remember the string.
  • This is what the manufacturer reported, not what the physician was paid in any sense they would recognize. The file totals every kind of general payment together; CMS's record for each physician lists the kind of each, and this file does not.
  • General, research and ownership payments are three different things and are never summed here. A research payment is money for a study, which may pass through an institution and reach the named physician in no part; CMS carries it in its own column for that reason.
  • A disputed transaction is one the recipient contested. CMS publishes it either way and counts it in the totals, so a total on this hub is a total of reported payments, some of which the recipient says are wrong.
  • A physician who appears with no payments in a program year is absent from that year rather than present with a zero. The file is payments, not a roster.
  • Covered_Recipient_Profile_ID identifies a CMS profile, not a person. CMS marks profiles it believes are the same person with Has_Multiple_IDs, and the supplement file carries the associations. A page that treats one profile as one doctor will understate anybody CMS holds twice.

Open Payments — payments grouped by covered recipient and nature of payment

Publisher
Centers for Medicare & Medicaid Services
License
U.S. Government work. CMS declares https://www.usa.gov/government-works as the license on every one of its 74 Open Payments datasets, with accessLevel: public. That page states a government work is "something created by a U.S. government officer or employee as part of their official duties" and is generally not protected by copyright, subject to conditions that do bind us: "You cannot use government materials in a way that implies endorsement by a government agency, official, or employee", and "You also cannot use federal government trademarks or federal government agency logos without permission."
Cadence
Annual
This copy fetched
Rows held
7,300,116
Columns
8 columnsprogram_year, profile_id, recipient_type, npi, nature_code, nature, transactions, amount
Pages built on it
Health

What is committed here is a summary. The full table has 7,300,116 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is d8452f60ecb23f995cbd54722ced7c946916271eb20f40d69571596988bb1ed4, so every figure here can be traced to the bytes it came from.

Every committed table for Open Payments — payments grouped by covered recipient and nature of payment, with its size and a download.
TableRowsSizeDownload
General payments by nature of payment and program year1052 KBCSV
Doctor-years compared with CMS's summary, by program year70 KBCSV

What it cannot show

— 4 limitations
  • General payments only. CMS's natures of payment describe general payments; research payments and ownership interests are separate records and are not in these files.
  • A nature is what the manufacturer reported the payment as. CMS does not audit the category, and a recipient can dispute a payment without disputing its nature.
  • The file carries a numeric code for each nature and no name for it. The name comes from CMS's state-by-nature file for the same year, paired by equal national totals, and a year whose codes do not pair exactly is not loaded.
  • Teaching hospitals are in CMS's file and are not loaded here: no page is built for one.

OPM FedScope Employment Summary Data

Publisher
U.S. Office of Personnel Management
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Quarterly, in March, June, September and December
This copy fetched
Rows held
14,511
Columns
11 columnsas_of_date, measure, agency_code, agency_name, parent_agency_code, location_code, location_name, occupation_code, occupation_name, employment, average_salary
Pages built on it
Federal spending
Download
Federal civilian employment by agency, state and occupation — gzipped CSV, 183 KB

What it cannot show

— 5 limitations
  • OPM does not publish every federal employee. Employees of the intelligence community and some Department of Defense positions are excluded, so a total here is the workforce OPM publishes and not the federal workforce. An agency absent from the file has not got no employees.
  • Average salary is an average over a group, not any person's pay. This is not a named-salary file and no figure here belongs to an individual.
  • Headcount is the number of employees in a status, not full-time equivalents, and it is a snapshot on the date of the release rather than an average over a period.
  • Rows of different measures describe the same employees from different angles. An agency row and a state row overlap, and summing across measures double-counts.
  • Where a measure has no agency, no location or no occupation, the column holds ALL rather than a blank, so that a row can be read without a join and an absent dimension is never mistaken for an unknown one.

Medicare Part D Prescribers — by provider

Publisher
Centers for Medicare & Medicaid Services
License
U.S. Government work. CMS declares https://www.usa.gov/government-works as the license on all three Part D Prescribers datasets, with accessLevel: public. That page states a government work is "something created by a U.S. government officer or employee as part of their official duties" and is generally not protected by copyright, subject to conditions that do bind us: "You cannot use government materials in a way that implies endorsement by a government agency, official, or employee", and "You also cannot use federal government trademarks or federal government agency logos without permission."
Cadence
Annual
This copy fetched
Rows held
1,416,883
Columns
39 columnsnpi, year, last_or_org_name, first_name, middle_initial, credentials, entity_code, city, state, zip5, ruca_description, prescriber_type, total_claims, total_30day_fills, total_drug_cost, total_day_supply, total_beneficiaries, ge65_suppression_flag, ge65_total_claims, ge65_total_drug_cost, ge65_beneficiary_suppression_flag, ge65_total_beneficiaries, brand_suppression_flag, brand_total_claims, brand_total_drug_cost, generic_suppression_flag, generic_total_claims, generic_total_drug_cost, opioid_total_claims, opioid_total_drug_cost, opioid_total_beneficiaries, opioid_prescriber_rate, opioid_la_total_claims, opioid_la_prescriber_rate, antibiotic_total_claims, antibiotic_total_drug_cost, antipsychotic_ge65_suppression_flag, antipsychotic_ge65_total_claims, antipsychotic_ge65_total_drug_cost
Pages built on it
Health

What is committed here is a summary. The full table has 1,416,883 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 9109b2b89dc1a227d5785279dde2f9ad6a46dac90a35f25efeea751b4ca0c6cf, so every figure here can be traced to the bytes it came from.

Every committed table for Medicare Part D Prescribers — by provider, with its size and a download.
TableRowsSizeDownload
Prescribing by state623 KBCSV
Prescribing by prescriber type2145 KBCSV
Prescribing nationally10 KBCSV
What CMS withholds, by measure90 KBCSV
The prescribers with the most claims1006 KBCSV
The prescribers with the highest drug cost1006 KBCSV
The prescribers with the most opioid claims1006 KBCSV

What it cannot show

— 6 limitations
  • This is Medicare Part D and nothing else. A prescriber's commercial, Medicaid and cash-pay volume is absent, so a small figure here can be a large practice. Nothing on a page built from this may be phrased as what a prescriber prescribes.
  • "To protect the privacy of Medicare beneficiaries, any aggregated records which are derived from 10 or fewer claims are excluded from the Part D Prescribers Dataset" — CMS, Part D Prescribers methodology, April 2026. So the file is not everything the prescriber prescribed, and CMS says in the same document that "summing data in detail file will underestimate the true Part D totals".
  • Separately from that exclusion, values are suppressed within a row: "beneficiary counts, claim counts, 30-day fill counts, drug costs, and day's supply are suppressed if the value is between 1 and 10 and also may be removed for counter-suppression purposes". A suppressed value is a blank, and CMS warns that "most statistical software packages will treat the 'blanks' as 'zeroes', resulting in underestimates of the true values". Nothing here coerces a blank to zero and no page prints one as zero.
  • Counter-suppression means a value can be withheld although it is above ten, because publishing it would reveal one that is not. A blank is therefore not evidence that a prescriber's count was small.
  • Prscrbr_Ent_Cd distinguishes an individual from an organization. An organization NPI is not a person, and this site's rule for pages about a person applies to one and not the other.
  • The opioid classification "is based upon opioid drugs included in the Medicare Part D Overutilization Monitoring System (OMS), which can change from year to year" — CMS, same document. An opioid measure is therefore not strictly comparable across years, and a trend drawn through it says so.

SBA Paycheck Protection Program (PPP) FOIA — loan-level data

Publisher
U.S. Small Business Administration
License
Public domain. SBA's catalog record for the dataset (DKAN metastore, identifier SBA-OCA-2022-07-001) gives its license as "https://www.usa.gov/publicdomain/label/1.0/", the usa.gov public-domain label, with accessLevel "public". The data is a federal agency's record of loans it guaranteed, a U.S. Government work under 17 U.S.C. 105. Read 24 September 2026 from https://data.sba.gov/api/1/metastore/schemas/dataset/items?show-reference-ids=false.
Cadence
Irregular (SBA's catalog says "irregular"; the program is closed)
This copy fetched
Rows held
11,468,210
Columns
55 columnsloan_number, date_approved, sba_office_code, processing_method, borrower_name, borrower_address, borrower_city, borrower_state, borrower_zip, loan_status_date, loan_status, term_months, sba_guaranty_percentage, initial_approval_amount, current_approval_amount, undisbursed_amount, franchise_name, servicing_lender_location_id, servicing_lender_name, servicing_lender_address, servicing_lender_city, servicing_lender_state, servicing_lender_zip, rural_urban_indicator, hubzone_indicator, lmi_indicator, business_age_description, project_city, project_county_name, project_state, project_zip, congressional_district, jobs_reported, naics_code, race, ethnicity, utilities_proceed, payroll_proceed, mortgage_interest_proceed, rent_proceed, refinance_eidl_proceed, health_care_proceed, debt_interest_proceed, business_type, originating_lender_location_id, originating_lender_name, originating_lender_city, originating_lender_state, gender, veteran, non_profit, forgiveness_amount, forgiveness_date, source_file, as_of
Pages built on it
PPP loans

What is committed here is a summary. The full table has 11,468,210 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from. The whole table is also a bulk download, below, as CSV and Parquet.

The full table’s SHA-256 is 1c12b62f7a8b9527aff303b827c4228767e405a5117e975c6514eb5a9a6eb3f9, so every figure here can be traced to the bytes it came from.

Bulk download: PPP loans, by borrower

Every row, SBA’s loans as of September 30, 2024, built September 26, 2026. Our file is licensed CC BY 4.0: credit “American Factbook (americanfactbook.org)”. The source is the publisher’s own files, u.S. Government work, public domain (17 U.S.C. 105).

PPP borrowers. One row per borrower, SBA's loans grouped by the name and address SBA printed, without the street address.

  • ppp-borrowers.csv.gz — gzip-compressed CSV, 484.9 MB, 10,688,106 rows
    SHA-256 bdce47770a6b19958f68141a711d34cf32cff7d5748e1fd3c97270fa9c284fcf
  • ppp-borrowers.parquet — Parquet, 406.1 MB, 10,688,106 rows
    SHA-256 75da14fba36920c351c25a73921701543444c3702ee5f39fd8c1ff711a5a412e
Columns — 10
borrower_id
— The borrower's key on this site, and the last part of the address of its borrower page. Loans are one borrower where SBA printed the same name and address.
borrower_name
— The borrower as SBA printed it.
city
— The borrower's city, as printed. The street address is left out.
state
— The borrower's state.
zip
— The borrower's ZIP code, as printed.
loans
— Loans SBA published for the borrower, first and second draw.
amount_approved
— SBA's current approval amount, summed over those loans.
amount_forgiven
— SBA's forgiveness amount, summed; empty where none is recorded.
first_approved
— The earliest approval date among the loans.
last_approved
— The latest.

PPP loans, by borrower. Every loan number with the borrower it was grouped under, so a grouping can be checked against SBA's files.

  • ppp-borrower-loans.csv.gz — gzip-compressed CSV, 265.0 MB, 11,467,506 rows
    SHA-256 b83b40fa918dfd67f3094081cedc420fad6df3f634f3b40fb206bf499924be6d
  • ppp-borrower-loans.parquet — Parquet, 246.9 MB, 11,467,506 rows
    SHA-256 92a309cd7249aec0823be3c449303dfd32b5c20338db1b73a56c21893d953488
Columns — 2
loan_number
— SBA's loan number, which finds the loan in SBA's own files.
borrower_id
— The borrower the loan was grouped under.

Checked before publishing: 10688106 borrowers equal the borrower count this site stores; 11467506 loans mapped, each once, equal SBA's 11468210 less 28 withheld and 676 placeholder names.

Every committed table for SBA Paycheck Protection Program (PPP) FOIA — loan-level data, with its size and a download.
TableRowsSizeDownload
Loans by business type262 KBCSV
Loans by county3,285176 KBCSV
Loans by industry1,26964 KBCSV
Loans by lender1229 KBCSV
Loans by month approved231 KBCSV
Loans by state594 KBCSV
Loans whose borrower name is not a name552 KBCSV
Loans nationally10 KBCSV

What it cannot show

— 8 limitations
  • A borrower has no identifier in the file. There is no EIN, UEI or DUNS number. A loan has one (LoanNumber) and a borrower does not, so two loans belong to one borrower here only where the name and address match exactly as published. A business that moved, or spelled its name two ways, is two borrowers.
  • Many borrowers are people. Sole proprietors, independent contractors and the self-employed borrowed under their own names, and 37% of one measured file is those three business types. A page prints the name, city, state and ZIP, and not the street address.
  • A loan is not a grant, and forgiveness is not a finding about the borrower. A charge-off means SBA paid the lender under its guarantee. The file carries no fraud flag, and nothing built on it may imply one.
  • JobsReported is the number of employees the borrower put on the application. It is not jobs saved, and SBA does not verify it in this file.
  • Race, ethnicity, gender and veteran status are self-reported and mostly unanswered: in the ≥$150k file, 768,556 of 968,524 loans have no race. A distribution of the answered minority is not a distribution of borrowers.
  • “Exemption 4” in the status column is FOIA's trade-secret exemption: SBA writes it where a loan is disbursed and neither paid in full nor charged off, and withholds the status. It is an open loan, not a status.
  • Twenty-eight loans have no location and no lender. In public_up_to_150k_1_240930.csv, 28 loans ($658,902.94; 26 paid in full, 2 charged off) carry a loan number, a borrower name, dates, amounts and a status, and every address, lender, office and project field blank. They are in every national total, and the state, county and lender tables count them as a row with no location rather than dropping them. Measured over all 11,468,210 loans on 24 September 2026.
  • The files are Windows-1252, and SBA's export replaced some characters with ¿ (“501(c)3 ¿ Non Profit”). Names are printed as published, ¿ included.

Treasury Fiscal Data

Publisher
U.S. Department of the Treasury, Bureau of the Fiscal Service
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Daily for the debt, monthly for the statements
This copy fetched
Rows held
97,779
Columns
8 columnsrecord_date, dataset, table_name, classification, parent_classification, measure, amount, unit
Pages built on it
Federal spending
Download
Treasury fiscal data: debt, budget and interest by month — gzipped CSV, 1.0 MB

What is committed here is a summary. The full table has 97,779 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is ff5f5006f39d85b8f0324072345b121f8c8fd5e576bd9f02f246a76038a75aef, so every figure here can be traced to the bytes it came from.

What it cannot show

— 5 limitations
  • These are national aggregates. There is no recipient, no state and no award here, and nothing on these pages can be attributed to a particular contract or grant.
  • Receipts and outlays are reported monthly and are revised. A figure is only as current as the snapshot it was drawn from.
  • Outlays reported by Treasury and obligations reported to USAspending are different measures collected under different rules. They do not reconcile, and a page that shows both says which is which rather than presenting the difference as an error.
  • The Monthly Treasury Statement reports a department's total on its own row. Agency subtotals include proprietary receipts, which are negative, so the parts of a department do not sum to its total.
  • Monthly Statement of the Public Debt amounts are published in millions. They are converted once, on the way in, and stored in dollars like every other figure here.

USAspending API v2

Publisher
U.S. Department of the Treasury, Bureau of the Fiscal Service
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Daily, with monthly agency submissions
This copy fetched
Rows held
8,351
Columns
9 columnsfiscal_year, measure, toptier_code, agency_name, agency_slug, award_type_group, state_code, state_name, obligated_amount
Pages built on it
Federal spending
Download
Obligations by agency, award type and fiscal year — gzipped CSV, 100 KB

What is committed here is a summary. The full table has 8,351 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is a7a02efe7eb6967db3b922aead44c699419aecc39b1809c3cb160f5dc6e3b6f5, so every figure here can be traced to the bytes it came from.

What it cannot show

— 9 limitations
  • USAspending reports obligations and outlays as agencies submitted them. It is not an audited financial statement and does not reconcile to one.
  • Award records are revised after publication. A figure is only as current as the snapshot it was drawn from.
  • Place of performance is the location the recipient reported, not necessarily where work was carried out.
  • An award amount is the value of the award, not money spent. Obligations and outlays are different measures and are not interchangeable.
  • Insurance figures are face value, like a loan guarantee, not money paid. The publisher's reporting of them changes sharply after FY2013, which is a change in reporting and not in spending.
  • The history starts at FY2008. The API states its own floor on every search response: "time period start and end dates are currently limited to an earliest date of 2007-10-01". Federal spending did not start there; our record of it does.
  • The open fiscal year is incomplete. It is not comparable to a closed year, and a fall against last year may only mean the year is not over.
  • The agency-by-state matrix covers the current fiscal year only. Agency totals and state totals cover every year.
  • An agency's slug is derived by USAspending from its name. An agency that is renamed gets a new slug, and its history appears under two identities rather than one.

USAspending Award Data Archive — contracts

Publisher
U.S. Department of the Treasury, Bureau of the Fiscal Service
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Monthly; each fiscal year's full file is republished under a dated name
This copy fetched
Rows held
2,404,831
Columns
18 columnsfiscal_year, part, recipient_key, recipient_uei, recipient_name, recipient_parent_uei, recipient_parent_name, agency_code, agency_name, sub_agency_code, sub_agency_name, action_type_code, action_month, state_code, actions, awards, recipients, obligations

What is committed here is a summary. The full table has 2,404,831 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is 2fa492959332ae8eb8520bd996116bb7fadb51fa155edebad3f0aecdcda64ae9, so every figure here can be traced to the bytes it came from.

Every committed table for USAspending Award Data Archive — contracts, with its size and a download.
TableRowsSizeDownload
The largest contractors and defense contractors, by fiscal year2,20042 KBCSV
Contract terminations by fiscal year, reason, agency, sub-agency, state and month2,62741 KBCSV
How many vendors and parent companies had a contract terminated220 KBCSV
The parent companies with the most contracts terminated1,10028 KBCSV
The vendors with the most contracts terminated1,10029 KBCSV

What it cannot show

— 7 limitations
  • USAspending reports contract actions as agencies submitted them to FPDS. It is not an audited financial statement and does not reconcile to one.
  • An obligation on a termination action is that modification's change to the contract, usually a de-obligation. It is not the contract's value, and it is not a saving.
  • Code F covers partial as well as complete terminations for convenience, so a contract counted as terminated may continue in part.
  • Awards are counted, not actions: one award can carry several termination actions. A General Services Administration schedule order is an award, so GSA's counts are mostly orders under schedules.
  • Before April 2022 contractors were identified by DUNS number rather than UEI. Where an action carries no UEI, the recipient is keyed by its DUNS number, and failing that by its name, so one company can appear under two identifiers across that change.
  • Rows of different parts describe the same actions at different grains and must never be summed together.
  • The open fiscal year is not held. Its file is republished as the year goes on, and a partial year is not comparable to a closed one.

USAspending API v2 — dimension aggregations

Publisher
U.S. Department of the Treasury, Bureau of the Fiscal Service
License
U.S. Government work, public domain (17 U.S.C. 105)
Cadence
Daily, with monthly agency submissions
This copy fetched
Rows held
210,485
Columns
11 columnsfiscal_year, measure, award_type_group, parent_code, code, name, slug, obligated_amount, outlayed_amount, award_count, population
Download
Obligations by recipient, place, program and more — gzipped CSV, 666 KB

What is committed here is a summary. The full table has 210,485 rows and is not in the repository — it would grow without limit as the publisher republishes. It is held in object storage and in the database, and the pages that show one row at a time read the database. What is committed, and downloadable below, is the aggregates and the ranked lists the published pages are built from.

The full table’s SHA-256 is efb21beb718044d4584b1ea1052265c8d47464ba9c6e483b85b7af9733ea6496, so every figure here can be traced to the bytes it came from.

What it cannot show

— 11 limitations
  • USAspending reports obligations as agencies submitted them. It is not an audited financial statement and does not reconcile to one.
  • Rows of different measures describe the same dollars from different angles. A recipient row and a county row overlap, and summing across measures double-counts.
  • The history starts at FY2008. The API states its own floor on every search response: "time period start and end dates are currently limited to an earliest date of 2007-10-01".
  • The open fiscal year is incomplete and is not comparable to a closed one.
  • Recipient rows are the largest recipients in a year, not every recipient. The publisher reports over eighteen million recipients across the whole record; we hold the top of each year and every page built on them says so.
  • Industry (NAICS) and product-or-service (PSC) codes describe contracts. They are requested with contract award types only, because a grant has no NAICS code.
  • Assistance listings (CFDA) describe financial assistance. They are requested with assistance award types only: the same request with contract award types returns an empty result set.
  • Account data — budgetary resources, outlays and budget functions — begins at FY2017 Q2, not FY2008. The publisher says so on every earlier request: "Account data powering this endpoint were first collected in FY2017 Q2 under the DATA Act; as such, there are no data available for prior fiscal years." Award data and account data therefore have different floors and a page that shows both states both.
  • Gross outlays are routinely negative at fine grain, so the word "spent" is never written over them.
  • Budgetary resources double-count transfers between accounts, by the publisher's own glossary, so a sum across accounts is not a total.
  • Place of performance is the location the recipient reported, not necessarily where work was carried out. A county figure is where the work was reported as performed, not where the money ended up.

Named but not yet held

These sources are part of a hub’s plan and have no copy here yet. They appear on the hub that will use them, and they will appear in the register above on the day they are first fetched.

How to cite

Cite the agency and the date of the copy you used, not this site. A figure taken from here on is a figure from the agency’s file as it stood on that date, and agencies revise their files. If you want the version we used rather than the version live today, the download above is that version.

A page for a single recipient, agency office, county, program or other spending entity is also served as JSON, exactly as the page is built: its address with /api/v1 in front, such as /api/v1/spending/recipient/id. No key is needed, and each address may make sixty requests a minute. The kinds it serves are listed in the OpenAPI description, and each such page links its own record from its sources.

The figures a hub’s pages were built from on a given day stay where they were. Each hub’s committed figures as they stood at the end of a day are at /data/as-of/YYYY-MM-DD/hub.json, for example /data/as-of/2026-09-26/spending.json. A day with no release of its own answers with the latest one before it. Those files hold the national, state and ranking tables. The pages for individual recipients, organizations and providers are read live from the database, and are not in them.

How these tables are produced, and what we do and do not change in them, is on the methodology page. Errors we have found and fixed are logged on corrections. If you find one, tell us.