How federal spending is controlled, and the Social Security and Medicare trust funds: how this investigation was made¶

Prepared October 5-6, 2026. Revision 1. Not yet independently reviewed.

1. What was asked and what was produced¶

The investigation, its pages and every file in this folder were produced by an AI assistant (Claude, running in the Claude Code tool) on instructions from the repository's owner, in one working session that ran from October 5 into October 6, 2026. The owner asked seven questions and then asked for everything said in answer to be collected into a govspends investigation "similar to what was done for Harris County", as the start of the website's United States federal government section. The owner supplied questions only, not documents, figures or conclusions. Nobody was contacted and no public-records request was filed.

#	Question (condensed)	Where the answer is now
1	Does the government track what it has spent against what the law authorizes?	Report Sections 1 and 3
2	How is multi-year and no-year money identified and tracked?	Section 2
3	How is mandatory spending (Social Security, Medicare, etc.) estimated and tracked?	Section 4
4	Are Medicare Parts B and D funded by premiums?	Section 5
5	Is Medicare Part A fully funded by its trust fund?	Section 6
6	How much general revenue goes to Parts B and D?	Section 5
7	What is the funding situation of OASI and DI?	Section 7
8	Collect all of it into the govspends repository and start the federal website	This repository

The original answers are kept unchanged in notes/qa_session_2026-10-05.md, apart from the condensing that note's own preamble describes.

2. How the evidence was gathered¶

Answers 1 to 4 were written from the assistant's general knowledge, with no documents consulted. Answers 5 to 7 used web searches, which turned up the 2026 Trustees Reports and secondary coverage (think-tank and news analyses). None of that secondary material is relied on in the report.
Sourcing. For the report, three research agents (sub-tasks of the same assistant session) were each given one topic and a list of every claim made in the answers on that topic:
trust funds;
budget execution and appropriations law;
mandatory spending, scoring and tax law.

Each agent was told to: - download government documents only; - save them in full under sources/official/<publisher>/ and log the URL, date and description in data/downloads_<topic>.tsv; - extract their text with scripts/extract_text.py; - record each verified fact in data/claims_<topic>.jsonl with the exact words of the document and its PDF page; - report every place where the documents contradicted the answers.

The assistant then wrote the report pages from the claims, adding claims of its own in data/claims_report.jsonl where a sentence needed a source not already recorded. 3. Checking. scripts/check_claims.py verifies every claim's quotation against the text of the cited page after normalizing whitespace, letter case, quotation marks and dashes. It also requires that every saved file have a download record and a text extraction. It recomputes the derived figures into data/derived.json and fails if a recomputed total disagrees with the printed one (HI surplus, OASDI net change, OASDI end-2025 reserves). The final run reported 452 claims, 16 derived figures and 0 failures. 4. Access problems. cbo.gov returned HTTP 403 to every scripted request. The five CBO documents are the unmodified original files from Internet Archive captures of their cbo.gov addresses; the capture addresses are in the download log. ssa.gov refused browser-style user agents but accepted a Wget user agent. uscode.house.gov was down for maintenance, so statutes come from GovInfo's 2024 edition of the U.S. Code.

3. Copied and computed figures¶

Every figure in the report is copied from a document, except the following, which are computed (data/derived.json):

the Parts B and D general-revenue total ($571.7 billion);
its share of SMI spending (74.7 percent);
its ratio to fiscal-year 2025 outlays (8.2 percent; this mixes calendar and fiscal years and is labeled as approximate);
the HI income shares in Table 6;
the Part B premium share (25.9 percent);
the Part D state share (10.4 percent);
the mandatory share of fiscal-year 2025 outlays (59.5 percent, and 73.3 percent with net interest; computed from CBO's figures, as neither CBO nor OMB states these shares for 2025).

4. Corrections to the original answers¶

These are the places where the government documents contradicted, refined or failed to support what the answers said. Items marked not in report could not be sourced to a government document and were left out of the report.

Question 1: spending against what the law authorizes¶

SF-133 frequency: the answer said "quarterly (monthly for large accounts)". The SF 133 is due at the end of each quarter plus November, July and August, for every open TAS (A-11 sec. 130).
SF-133 publication: that SF 133s are "published on OMB's MAX portal" is not in report.
Who apportions: the statute gives apportionment to the President (31 U.S.C. 1513(b)); OMB does it by delegated executive order.
Penalties: penalties for exceeding an apportionment are in 31 U.S.C. 1518-1519. Sections 1349-1350 apply to 1341 and 1342.
Public apportionments: sourced in detail (P.L. 117-103 sec. 204(b); made permanent by P.L. 117-328). The site was removed on or about March 24, 2025, and ordered restored by the D.D.C. ruling of July 21, 2025 and the D.C. Circuit order of August 9, 2025. The restoration date itself is not in report.
Impoundment decisions: "several decisions in 2025" became 11 decisions, 7 of which found at least one violation.
USAspending File A: File A is not taken from GTAS. It comes from agency financial systems and is validated against GTAS.
Expired authorizations: "several hundred billion dollars" became $494 billion in FY2026 (CBO, 2026).
Size of mandatory spending: "about two-thirds of outlays" was inaccurate. Mandatory spending was 59.5 percent of FY2025 outlays, or 73.3 percent including net interest.
USAspending data quality: "GAO repeatedly flags gaps in USAspending data" is not in report.

Question 2: multi-year and no-year money¶

Example account: the example account 075 ... 0512 was illustrative, not a real account. The report uses SSA's account 028-8704, which has real annual, two-year and no-year TAS.
TAS components: current Treasury and USAspending definitions list seven components (ATA, AID, BPOA, EPOA, availability type code, MAIN, SUB). Availability type codes are X (no-year), F (clearing/suspense) and C (default).
Trust fund account numbers: "8000-8999 are trust funds" is refined. Non-revolving trust fund accounts are 8000-8399 and 8500-8999, and trust revolving funds are 8400-8499 (A-11 sec. 20.11).
Health care trust fund payment account: the Payments to Health Care Trust Funds account (075-0580) is annual, not no-year.

Question 3: mandatory spending¶

Depletion dates: the 2025 Trustees Reports' dates are superseded by the 2026 reports: OASI fourth quarter of 2032, combined OASDI third quarter of 2034, HI second quarter of 2033.
Benefits payable after depletion: "roughly 77-81 percent" became 78 percent (OASI) and 83 percent (OASDI).
Antideficiency Act and depletion: that the Antideficiency Act is what limits payments after depletion is not in report. The report cites the Financial Report instead: the program "lacks the authority to borrow".
Medicare 2% sequester: "in effect continuously since 2013" was wrong. It was suspended from May 1, 2020 to March 31, 2022, and was 1 percent from April to June 2022.
Baseline rule: the $50 million rule applies to programs established by laws enacted on or before August 5, 1997. Separately, any program over $50 million under an expiring law is assumed to continue.
PAYGO: the scorecards were set to zero by P.L. 118-158 and then P.L. 119-37, and a PAYGO sequester has never been triggered.
Interest appropriation: the permanent appropriation for interest is 31 U.S.C. 1305(2).

Question 4: Parts B and D premiums¶

Part D financing shares: the answer gave about 74 percent general revenue, 15 percent premiums and 11 percent states. For 2025 the figures are 81.2, 8.1 and 10.4 percent. This was corrected in the original answer to Question 6.
Part D base premium: the 2026 base premium ($38.99) was set by the Inflation Reduction Act's 6 percent growth cap, not the 25.5 percent formula, which would have given $75.38.
Account 075-0580: describing Payments to the Health Care Trust Funds as an "appropriated entitlement" could not be sourced. The account is appropriated annually ($593.8 billion in P.L. 119-75, plus "such sums as may be necessary" for unanticipated payments). The report describes it that way and does not apply the label.
Funding warning frequency: "issued almost every year since 2007" became warnings in the 2007-2013 and 2018-2026 reports.
Funding warning trigger: the trigger is precise. Spending minus dedicated financing must exceed 45 percent of spending within seven years, and a warning follows two consecutive determinations.
Congressional response: "Congress has never acted on one" is not in report.
Premiums paid by Medicaid: the share of premiums paid by Medicaid (Medicare Savings Programs) is not quantified in any saved document and is not in report as a figure.

Question 5: Part A¶

Cash flow: "HI has spent more than it takes in during several recent years" was wrong for 2025, when HI ran an $18.2 billion surplus. Deficits are projected from 2027.
HI income shares: these became exact: 87.2 / 8.9 / 2.0 / 1.3 / 0.2 / 0.4 percent.
75-year deficit: the 75-year HI deficit is 0.56 percent of taxable payroll in the 2026 report, not "about 0.4" (0.42 in 2025).
Part A premiums: "people who lack 40 quarters" pay a premium is refined. The full premium ($565) applies below 30 quarters, and a reduced premium ($311) applies to 30-39 quarters.
Additional Medicare Tax: the 0.9 percent tax is paid by the employee only. There is no employer match.
Unverified details: the Trustees Report table number "III.B.4" and the Medicare Advantage split between HI and SMI are not in report.

Question 6: general revenue for Parts B and D¶

Share of GDP: "2.1 percent of GDP in 2026 rising to 3.9 percent by 2100" became 1.9 percent in 2025 rising to 3.7 percent in 2100.
Consecutive-year count: the "10th straight year" mixes two counts. 2026 is the 10th consecutive determination and the 9th consecutive warning.
Share of all Medicare: "roughly half of all Medicare spending" became 47 percent of Medicare's non-interest income.
Comparison with other programs: "more than any other single program except interest" is not in report.

Question 7: OASI and DI¶

Actuarial deficit decomposition: the table taken from secondary coverage was partly wrong. It omitted the economic (+0.10), disability (+0.02) and methods (-0.05) changes, and gave fertility as -0.35 instead of -0.33. The demographic total is -0.44.
Size of the jump: "the biggest one-year jump since 1994" came from secondary coverage and is not in report.
Covered workers: the Trustees Report says 185 million; CRS says about 186 million. The report gives both.
Railroad Retirement interchange: combined cost includes $6.0 billion of financial interchange with Railroad Retirement, which the answer omitted.
DI in 2016: that DI "was close to running out in 2016" is not in report. The 2016-2018 reallocation under P.L. 114-74 is.
2025 tax law: the Trustees Report names the One Big Beautiful Bill Act but not its public-law number. The report names the Act only.

5. Judgment calls¶

Placement: the investigation is filed under the national government (us/national) as one investigation with seven sections, rather than seven investigations. The questions build on each other.
What counts as a source: only government documents count. CRS reports are treated as government documents; think-tank and news analyses used in answers 6 and 7 are not cited.
CBO archive copies: CBO documents retrieved through the Internet Archive are accepted as originals because they are the raw captured files.
Calendar against fiscal years: where calendar-year Trustees figures are set against fiscal-year budget totals, the report says so.

6. Problems found during the work¶

Download logs: two research agents shared a scratch directory. Some of the mandatory-spending agent's download records were first written into the execution agent's log. That agent moved them to its own log; check_claims.py confirms every file has exactly one record.
Transcription: the first transcription of answer 1 into notes/qa_session_2026-10-05.md was an earlier draft wording rather than the answer as given. It was replaced with the exact text before anything was published.
Statements with no source: during writing, several sentences in the report were found to say more than their sources. They were cut or reworded, and the cases are listed in Section 4. Examples include Medicaid paying Part B premiums and account 0580 as an appropriated entitlement.
Third-party credential in saved pages: three CMS web pages (cms/2026_part_ab_premiums_fact_sheet.html, cms/2026_partd_bid_premium_stabilization_fact_sheet.html, cms/fy2025_improper_payments_fact_sheet.html) embed a Mapbox access token in a site-wide settings script. GitHub's secret scanning refused the first push because of it. The token's value was replaced with a [REDACTED: ...] marker on that one line of each file. No visible text changed and no claim relies on that line, but these three files are no longer byte-for-byte what cms.gov served. Their download records say so.

7. How to re-verify¶

From this folder:

python3 scripts/check_claims.py        # every quotation on its page; derived figures; download records
python3 scripts/build_references.py --check
sha256sum -c MANIFEST_SHA256.txt
PDFTOTEXT=$(which pdftotext) python3 scripts/extract_text.py   # optional: rebuild the text extractions (needs poppler-utils)

To check one figure, find it in data/claims_*.jsonl and open the source file at the stated PDF page.