Skip to content
Fiscal Receipts

Research tools / Reproduce

Data downloads

Take the underlying files into your own research workflow. Each export includes its schema; the citation index connects recorded facts back to their sources.

The snapshot includes 143,804 source citations and 1,938 programs in the FY2026 index. Datasets marked “cited” include row-level citation linkage; others are citation-tier pending (see methodology).

Schema & data dictionary · source linkage

Schemas & data dictionary: every Parquet file embeds its column schema (readable via DuckDB DESCRIBE), and the data explorer lists each dataset with row counts and a description of what one row represents. Keep the citation index alongside your data export so fact IDs remain traceable after a join or calculation.

Bundle built: . Files are in Apache Parquet format, readable with DuckDB, pandas, R arrow/duckdb, or any Parquet-compatible tool. Each file includes the same provenance metadata that backs on-screen figures.

budget_lines8,553 rowscited

One row per (program element × amount type) figure in the FY2026 President's Budget R-1/P-1 workbooks, carrying the source sheet and cell address it was read from. Fenced to the PB2026 edition. Add rows only: the P-1 flags each row Add or Non-Add, and Non-Add rows are memo lines the exhibit does not add into its own totals, so summing the workbook without that filter double-counts. P-1R rows are the reserve-component subset of P-1 lines and are never summed with them.

budget_lines.parquet
budget_lines_decade56,294 rowscited

The decade sibling of budget_lines: one row per (President's Budget edition × program element × amount type) figure across all ten editions PB2017–PB2026, each cited to its own edition's workbook cell (a PB2017–PB2023 P-1 row's pe_bli is its era_key; p1_era_line_map gives the code printed on it). Editions are parallel publications, never reconciled.

budget_lines_decade.parquet
p1_era_line_map6,927 rows

One row per era-keyed PB2017–PB2023 P-1 display line (the Classified Programs lines, loaded under 9999999999, are not in the map): era_key (its pe_bli in budget_lines_decade), line_item_code (the budget line code printed on it), program_key (the pe_bli a same_program line's era citations carry; the printed code except 0390DOM for 0390D's O&M line, whose "&" cannot appear in a web address) and the dated owner decision that joins it to a program page or to history only, or excludes it. Only a same_program line whose code has a program page, or an F-15 family-history line, has those rows and citations. Every row is a decision; none carries an amount.

p1_era_line_map.parquet

Procurement lines before PB2024

Each PB2017–PB2023 P-1 line reaches a program page only through a dated decision on the budget line code it printed, published in the p1_era_line_map dataset above. For each edition the table counts the lines, the codes each decision covers and the era figures cited on the site that carry a complete PDF receipt.

Era P-1 lines, codes per decision and complete PDF receipts, by President's Budget edition.
PB2017Lines969Same program697 codesHistory only209 codesExcluded4 codesPDF receipts2,208 of 2,235
PB2018Lines1,029Same program730 codesHistory only208 codesExcluded4 codesPDF receipts2,325 of 2,358
PB2019Lines989Same program746 codesHistory only155 codesExcluded4 codesPDF receipts2,371 of 2,406
PB2020Lines975Same program780 codesHistory only136 codesExcluded3 codesPDF receipts2,468 of 2,499
PB2021Lines1,005Same program810 codesHistory only119 codesExcluded4 codesPDF receipts2,556 of 2,625
PB2022Lines993Same program840 codesHistory only86 codesExcluded2 codesPDF receipts2,673 of 2,700
PB2023Lines967Same program857 codesHistory only61 codesExcluded1 codePDF receipts2,661 of 2,700
jbook_details21,993 rowscited

One row per (program element × project × budget scenario) cost figure extracted from J-book R-2/P-40 XML, with its XML element path and source-PDF SHA-256. `account` is the appropriation the figure was filed under (NULL for R-2/RDT&E rows, which carry none): ten PB2026 budget-line codes are shared by two different Navy programs in two different appropriations, so summing this table by pe_bli alone adds those pairs together.

Page resolution is partial: of 17,900 rows carrying a non-zero amount, 3,407 resolve to a unique PDF page and 6,472 to the first page the amount appears on; the remaining 8,021 (45%) resolve to no page and cite their XML path instead.

jbook_details.parquet
jbook_narratives11,717 rowscited

One row per J-book narrative text block (mission, description, justification or accomplishment/planned-program) with its XML element path and source-PDF SHA-256. Fenced to the PB2026 edition, plus the PB2017–PB2025 narratives a program-lineage edge cites: citation targets only, never a program page's own prose.

jbook_narratives.parquet
fct_budget_trajectory1,989 rowscited

One row per (program element × organization) with FY2024 actuals, FY2025 total, FY2026 total and the FY25→FY26 delta pivoted side by side.

fct_budget_trajectory.parquet
dim_programs1,936 rowscited

One row per program element that has full R-2/P-40 J-book detail (the detail-grade tier, NOT the full page universe; see the corpus statement on /data/), with org, exhibit family, project count and reconciliation status.

dim_programs.parquet
dim_entities116,442 rowscited

One row per contractor entity family, name-normalized across USAspending/SAM.gov UEI registrations, with its UEI count, total DoD obligation and worst match-confidence tier. The sam_* columns carry the SAM.gov Entity Management record of the family's dominant registration (status, CAGE, legal business name, business types, primary NAICS, expiry) where the bounded extract has reached it — enrichment, never an input to the confidence tier, and NULL wherever it has not.

dim_entities.parquet
fct_influence210 rowscited

One row per (contractor family × filing year) of Senate LDA lobbying totals, beside the family's DoD obligations (repeated on each year row). Income and expense are non-additive: a self-filer's expense can include its outside firms' income, so lobbying_total_usd, a plain sum, can double-count.

fct_influence.parquet
fct_program_lobbying12,571 rowscited

One row per (LDA filing × matched program element) mention — a filing appears once for every program its issue text has keyword co-occurrence with (an exact PE/BLI code, a curated alias, or >=2 distinct title words; see evidence_kind), never a claim that the filing names the program. Rows count evidence-tiered mentions, not filings.

fct_program_lobbying.parquet
fct_budget_to_awards3,634 rowscited

One row per (program element × USAspending award) crosswalk link, each carrying its match method and confidence tier. A link is an inference, not a reported fact.

fct_budget_to_awards.parquet
dim_geography1,048 rowscited

One row per (pop_state × pop_district) place of performance, with transaction count and total obligation over every DoD award transaction the site loads that records a pop_state, not only crosswalk-linked awards. pop_state is not normalized (contracts: two-letter code; assistance: state name).

dim_geography.parquet
fct_district_totals190 rowscited

One row per (state × congressional district), with obligation dollars counted once per DISTINCT high-confidence-crosswalked award — the district headline. Its per-(district, program element, appropriation account) sibling, fct_district_programs, is not summable: an award matched to N program elements appears N times with the same dollars there.

fct_district_totals.parquet
fct_program_concentration535 rowscited

One row per program element carrying at least one published crosswalk link, on TWO bases: *_all over every published link (high and medium confidence), *_high over high-confidence links alone. hhi_high and top_family_high are NULL below the floor: 3 linked awards, 2 families holding positive dollars (positive_family_count_high), positive program_dollars_high (ROADMAP #80).

7 of its 535 rows are code-level (scope = 'code'): a budget-line code two or more programs share. On 3 of them (0145, 3010 and 3215) more than one member's key carries links, so the row pools every member key with links and describes no single program. On 4 (2101, 2292, 3050 and 4217) one member's key carries every link, so the row is that member's figure: 2101 is 2101-WPN's, 2292 is 2292-WPN's, 3050 is 3050-OPN's, 4217 is 4217-OPN's. member_keys_with_links counts the member programs whose own key carries a published link.

fct_program_concentration.parquet
fct_state_per_capita6 rowscited

One row per (jurisdiction × comparable spending category × fiscal year) in the CA/CT state-checkbook pilot, with population and per-capita amount.

fct_state_per_capita.parquet
fct_improper_exposure15 rowscited

One row per federal agency reporting improper payments, with the derived dollar exposure and weighted error rate from paymentaccuracy.gov.

fct_improper_exposure.parquet
dim_lobbyists1,612 rowscited

One row per individual lobbyist named in Senate LDA filings, with covered-position and revolving-door flags and the UUID/URL of the filing that disclosed them.

dim_lobbyists.parquet
citations143,804 rowscited

One row per source citation, keyed by fact_id, in 10 kinds: workbook (a President's Budget workbook cell), derived (a recorded value with its formula), lda_filing (a Senate LDA filing), jbook_narrative (a J-book narrative passage), jbook_pdf (a figure printed in a J-book PDF, or the PDF alone where that figure is unresolved), usaspending (a USAspending API query), announcement (a DoD contract announcement), subaward (a subaward on a USAspending prime award), state_file (a state spending source file) and state_soql (a state open-data query).

citations.parquet

Additional assets (not in table above)

  • pdfs/ — SHA-named J-book PDFs
  • workbooks/ — 30 R-1/P-1 Excel rollup files