Methodology

Federal awards — USAspending.gov

The official federal award database (contracts, grants, loans, and subawards), mandated by the DATA Act. We download bulk ZIP files from files.usaspending.gov, convert them to compressed Parquet, and record the exact file name, URL, and SHA-256 hash of every file. Current scope: Department of Defense agencies, FY2017 onward. Source cadence: monthly — its least recently refreshed dataset (subawards) was last fetched 2026-06-11.

DoD budget justification books (“J-books”)

The detailed budget submissions the Pentagon sends to Congress each spring, published at comptroller.defense.gov. These cover every R&D program (R-2 exhibits) and procurement budget line (P-40 exhibits) with program narratives, project-level cost tables, and congressional justifications.

The Pentagon's budget system embeds its own database inside these PDFs: the full structured XML is attached inside the PDF itself. We extract that XML directly. For the rare document that lacks an XML attachment, we use a deterministic PDF parser as a fallback, with an accuracy gate requiring ≥98% numeric-field agreement against XML-backed ground truth. We also download the official R-1 and P-1 Excel rollups, which serve as independent control totals for total obligational authority (TOA). Source cadence: annual.

The P-1 flags each row Add or Non-Add, and we load the Add rows only: Non-Add rows are memo lines the exhibit excludes from its own totals, so summing the workbook without that filter double-counts.

Improper-payment estimates — paymentaccuracy.gov

Agencies are legally required to estimate and report payment-error rates. The dollar exposure figure we show per agency is derived by multiplying the published rate by the published outlay figure. The federal government reported approximately $186 billion in improper payments in FY2025. This is a derived estimate and is labeled as such. Source cadence: annual.

GAO high-risk list — gao.gov/high-risk-list

The GAO's biennial list of federal programs at high risk for fraud, waste, or mismanagement. This list is a department-level designation, and program pages label it as one. Source cadence: biennial.

Senate lobbying disclosures — lda.senate.gov

The Senate Lobbying Disclosure Act database (lda.senate.gov/api/v1) contains filings, each with a permanent UUID, registrant, client company, dollar amounts, agencies lobbied, and issue text. We have linked LDA client names to our company-family database and match each filing’s issue text against our program titles for keyword co-occurrence — never a claim that the filing names the program. 12,571 program mentions connect filings to budget lines this way: one row per filing × matched program element, qualifying only when the program’s PE/BLI code appears (an all-digit one only beside a budget-line label; none of the 1,631 bare-number matches counted before this rule had one), a curated alias appears, or at least two distinct, non-generic words of one program’s title co-occur in the same filing — a single common word is never treated as evidence (see the dataset inventory, which states the qualifying tier per row as evidence_kind). Lobbying income and expenditure by year are shown alongside federal obligations received — influence is presented side by side with outcomes, never as a causal claim. In those yearly figures an amendment replaces its original instead of adding to it (a correction decided 2026-09-26); where a quarter’s amendments disagree, the smallest counts (our copy keeps no posting date). Program mentions still count superseded reports.

State checkbooks — California and Connecticut (pilot)

California's Open Fi$Cal and Connecticut's OpenCheckbook publish transaction-level government spending. We aggregate by department, spending category, and fiscal year, and use Census Bureau population estimates (NST-EST series) for per-capita comparisons. The category mappings that bridge both states' classification systems are published alongside the data.

California data covers FY2025 only — CA Open Fi$Cal updates on a lag; prior years not yet ingested. why FY2025 only? →

Reconciliation

We apply two arithmetic checks to J-book figures. Check A: project-level amounts within an exhibit must sum to the program-element total in that same exhibit (tolerance: ±$0.001M). Check B: that program-element total must match the corresponding row in the official R-1 or P-1 Excel rollup. A failure does not suppress the figure — it marks the row unreconciled, files a review-queue record against the specific failed check, and the program page carries a Partial Reconciliation badge instead of a Reconciled one. A build gate then refuses any unreconciled row that has no queue record, so a figure can fail these checks in the open but never in silence. Both checks cover prior-year actuals, current-year enacted, and the budget-year request and its base; the J-book’s cumulative All Prior Years element has no R-1 or P-1 column to compare against, so it is never checked and never held against a line.

Zero-absent rule

When a program has no funding for a given fiscal year, the official rollup workbook simply omits the row. Our reconciliation code recognizes this so zero-funded programs are not incorrectly flagged as failures.

Coverage gate

At least 99% of R-1 program elements must have either extracted detail or an explicit gap record. Silent holes are a build failure.

Provenance spot-check

Each build randomly samples 50 served facts and mechanically verifies that the cited source document exists on disk, its SHA-256 matches the download manifest, and the XML element path resolves to a real node. A number whose citation chain breaks does not render.

Per-build automated checks

The automated check inventory includes Python and browser tests, alongside 27 site verification gates and 163 dbt data-model assertions. The analyst-agent evaluation set (48 question-answer pairs) requires at least 44 correct answers and 100% citation resolution.

Citation tiers

Every rendered figure is in one of three states. Cited (underlined, clickable): a fact_id resolves to a source document in our citation index — J-book PDF page-and-bbox, a specific workbook cell, an LDA filing, a USAspending query, or a derived formula whose inputs chain back to those sources. XML-path chip: a zero-dollar budget line that exists in the structured XML but has no corresponding citation row — the XML element path is displayed. Citation tier pending (⁂): figure is from a dataset for which row-level citation linkage is not yet complete. As of this build every published dataset carries a citation tier (the pending ledger is empty); the state remains defined — and gate-enforced — for future datasets that ship before their citations do.

Renames and acquisitions — hand-curated, not inferred

Name inference cannot know that Raytheon Company and RTX Corp are one company: federal award records carry the recipient name that was on the contract, and a rename produces two names. We close that gap with a small hand-curated table of corporate renames and acquisitions, each row sourced to an SEC filing or an official company press release. Those sources are external references, not warehouse citations — they leave this site and carry no fact ID, because we read them rather than extracted them. The combined figure a merged family renders IS a warehouse citation: a derived fact whose inputs are the member figures it replaced, recomputed by the citation verifier so a merge can never double count. Names that do not resolve to a recipient family are marked unresolved rather than guessed at.

Budget-to-contract links

Connecting a budget program element to the contracts that funded it is an inference. As of 2026-09-27, 9,588 of the 12,866 links the crosswalk grades high or medium carry a per-award hand adjudication — the most recent made on 2026-09-01 — recording which program elements, if any, the award's own contract record supports. 8,475 of those found work that could not be pinned to any one program element; 56 of them still publish, at medium, and the rest no longer publish. The fpds-ap and subaward+lexicon paths carry no per-link adjudication — their precision is sampled instead (below). High: affirmative program-level evidence — the contract names a program the budget line's own J-book pages also name. 60 of the 1,106 links published at high carry a per-award hand adjudication, all 60 challenged by two independent adversarial reviewers; the other 1,046 rest on the announcement+lexicon path. All 1,106 carry a recorded review, counted once under the strongest: 60 a two-lens hand adjudication, 417 a per-proposal verdict pair and 629 only a survivor-list entry, which records survival of the adversarial pass, not its verdict. Medium: weaker evidence, of more than one kind. An account-based link, where the award drew from the same appropriation account as the program, is an association, not evidence that this specific program paid for the contract. A held-out sample of account / sub-agency links, judged on program attribution, confirmed 0 of 58 (2026-09-11). On that figure the tier was withdrawn (decided 2026-09-25): its 8,865 links that no two-lens hand adjudication pinned no longer publish. Where the evidence is instead an FPDS acquisition-program tag or a subaward description (both below), the program is established but which of its budget lines paid is not. 91 more were demoted from high: 63 unadjudicated keyword matches, 5 refuted in review, 12 rejected by a reviewer and 11 refuted by the precision study. Low: only the account matches — never published. Our earlier automated high tier (account match plus keyword overlap) measured 9.1% precise under this adjudication (37 of 408 links confirmed) and was corrected on 2026-09-01; superseded links are retained in the correction record.

A second evidence path covers major acquisition programs: some DoD contract records carry an FPDS “Program, System, or Equipment” tag naming the acquisition program (F-35, Virginia class, Sentinel). We hand-mapped every such program to its J-book budget lines — each mapping challenged by a two-reviewer adversarial process — then linked a tagged award to a specific line only when the award's own funding accounts match that line's appropriation. FPDS-tagged awards publish at medium — the tag plus a verified program mapping establish the program, and the award's funding accounts confirm the money color, but which of a program's several lines (production vs. modification vs. research) paid is not provable from account data alone. A held-out study measured the earlier “unique line” high tier at 34 of 60 and it was withdrawn on 2026-09-04. Tagged awards whose funding is entirely outside the program's J-book accounts (for example O&M sustainment money) are not linked. The FPDS tag is DoD-entered and sparse — it appears on well under 1% of awards, concentrated in the largest programs — so absence of a link never means absence of spending.

A third evidence path reads the Department of Defense's own daily contract announcements (defense.gov, archived with snapshot timestamps and file hashes): each announcement names the contract number and describes the work, often by program. Where an announcement's program name is one that a program element's J-book narrative itself owns, the pair is a candidate; every candidate is judged by an agent reviewer and challenged by an independent adversarial reviewer, and only links surviving both publish. One stays at high only while a recorded review upholds it and no recorded rejection or refutation applies; otherwise medium, with the reason recorded. At high, the announcement establishes the contract, and the program is identified by its name as written, by a normalized designator, by an alias an adversarial reviewer checked, or, rarely, by the announcement's own description of the work. Where the adjudication packet recorded which of those applied, the link's citation card states it; where it did not, the card says the basis was not recorded rather than asserting one; an unrecorded basis is not evidence the announcement named the program outright. Announcement links additionally require the award's funding accounts to match the line's appropriation; awards funded only from operations and maintenance money are not linked to research or procurement lines. Names that merely describe a platform the contract supports, generic services, or weak generic names are rejected by design. Each link cites the announcement it came from.

Scope of the announcement path: deterministic name matching covered all 32,852 archived announcement records that join the award lake; 4,508 named a program a J-book narrative owns. Of the 28,344 that did not, 15,604 have been through an LLM-assisted alias pass — 89.1% of the residue by announced value; the remaining 12,740 records (10.9% of that value) were not attempted. Its most recent round carries its own measurement: two independent reviewers judged a random draw of 60 of its links, recorded 2026-09-19, and confirmed 48 of the 55 that publish — the others no longer publish and count in neither direction; those links only, not the path as a whole.

Where the only evidence is a subaward: FSRS subaward reports describe the work a subcontractor performs under a prime contract, and when that description names a program the PE's own narrative owns, the prime is linked at medium — the evidence is one hop removed, so it never publishes as high and its rationale names the subaward it rests on.

Measured precision of the published tiers, from a held-out hand-adjudicated sample re-run through a two-reviewer process and recorded 2026-09-04. Every figure answers one question — program attribution: does this award execute this program element? — not whether the linking rule fired. Each sampled link is counted under the tier it publishes under today, not the tier it carried when it was drawn; one the corpus no longer publishes is counted in neither direction: announcement+lexicon 57/61; fpds-ap 89/114; subaward+lexicon 52/59. The announcement tier's figure comes from the 2026-09-04 draw, which predates 367 of the links now publishing under that tier: the LLM-alias pass's most recent round added them, and the round's own held-out sample above measures its links instead. No precision figure is published for the remaining tiers a reader can meet — account, account+subagency, account+tokens. Those rest on an appropriation-account match, narrowed by a hand adjudication of the award, by sub-agency and a pinning two-lens adjudication, or by keyword overlap; how often that association names the right program has not been independently measured for these tiers.

Derived figures are labeled derived

Any figure computed from published rates or published subtotals — rather than directly reported in a source document — is labeled as derived wherever it appears.

Contractor concentration is computed on two bases and published on one

Both bases (high-confidence links alone, and every published link) share one crosswalk and ship in the downloadable warehouse with derived citations; a program page shows only the high-confidence figures, which it and the warehouse carry only over at least three such awards across two or more contractor families holding positive obligations with positive net linked dollars; below that floor it states the absence rather than substituting the wider figure. In the warehouse a shared code is one row, labelled scope = 'code'; /downloads/ says of each whether it is one program's figure.

Year-over-Year Swings (yoy_swing)

Programs where FY2025 total is ≥ $50M and the absolute percentage change to FY2026 is ≥ 50%. Budget figures come from the fct_budget_trajectory mart (trajectory pivot of FY2025 enacted and FY2026 requested budget workbook lines). The figure shown is the percentage change; the delta in dollar terms is the cited trajectory figure.

Zeroed in FY2026 (zeroed_fy2026) — withdrawn, currently empty

Programs the budget workbooks record as literally zero in FY2026 after carrying positive FY2025 money. The figure shown is the last known FY2025 amount. No FY2025 floor — any positive amount qualifies. This section currently shows nothing, and that is the correct result.

What we got wrong. Until August 2026 this section published 87 cards reading “<program> zeroed out in FY2026 (had $0 in FY25)”. Both halves of that sentence were false. The underlying query treated a program’s absence from our FY2026 extract as a zero, and none of the 87 was a genuine zero. Separately, a formatting bug printed the FY2025 amount as “$0” on every card, hiding $7.07B of real FY2025 money.

Why absence is not zero. In DoD’s published R-1/P-1 workbooks a blank FY2026 cell and a 0 cell mean different things, and both appear in the same file. Our loader preserves that distinction. A blank usually means the line is not carried in the FY2026 columns — most often because the program element was renumbered, not cancelled. DARPA is the clearest case: the FY2026 request retired 14 of its old program elements and introduced 8 new ones, while DARPA’s total request rose to $4.92B from $4.15B. Calling Defense Research Sciences “zeroed out” described a renumbering as a cancellation.

The query now requires positive evidence: the corpus must hold a FY2026 figure for the program and that figure must be zero. On the current corpus nothing meets that bar, so the section is empty rather than populated with inference. A gate (24h) fails the build if any feed card claims a program was zeroed in a year for which we hold no figure, or a non-zero one.

Award Concentration Shifts (concentration_shift)

Each program-year with at least $5M in positive obligations on high-confidence links, and its Herfindahl-Hirschman Index (HHI) over them; no HHI floor applies. On a budget line more than one program uses, only a program carrying every crosswalk link (high or medium) gets a card, linked to its page. HHI = sum(share² × 10,000) where share = family_obligation / total_obligation; only positive obligations are included. Bands follow the 2023 Merger Guidelines of the Justice Department and the FTC: an HHI from 1,000 to 1,800 is moderately concentrated and one above 1,800 is highly concentrated. Below 1,000 this site says unconcentrated, its own label for a range those bands leave unnamed.

This figure is a single fiscal year’s HHI, not a program’s overall concentration. A program’s own page renders a pooled HHI computed across every award year, on the basis and above the floor section 4 states. The two are legitimately different measures: a concentrated year can sit next to an unconcentrated pooled figure. Every card states which fiscal year its HHI covers and that the pooled figure can differ.

Largest Request-vs-Actuals Gaps (request_vs_actuals_gap)

The largest gaps between what a President's Budget asked for a fiscal year and what a later book reported actually spent: the PB(N) request for FY N compared against the PB(N+2) book's FY N actuals, ranked by absolute dollar gap across all loaded editions (PB2017–PB2026). The claim is scoped exactly to this request-vs-actuals comparison — the feed makes no claim about request-vs-request changes between adjacent asks. Each figure cites its book-diff derived fact; the citation panel shows the two input figures (each cited to its own edition's workbook) and the show-your-work breakdown.

New Defense Contractors (new_entrant)

Entity families whose first award year in the DoD transaction data is FY2024 or later and whose cumulative positive obligations exceed $1M. “First award year” is determined from the USAspending award archive (FY2017 onward). Entities that received their first award before FY2017 may appear as new entrants due to archive coverage limits — treat as a weak signal. The figure shown is total cumulative obligations (USD).