Methods
Where the figures come from, and how the model is checked
The data pages show official figures. The policy lab is a model built here. This page says how each is made and tests the model against the Congressional Budget Office, lever by lever.
The data pages
Every figure on the data pages is read from a federal file: the White House budget office's historical tables, Treasury's daily debt and monthly statements, USAspending, the IRS, the Census Bureau, Social Security, Medicare and Medicaid, Veterans Affairs, the Agriculture Department and PaymentAccuracy.gov. A script downloads them, and the page writes its sentences from the same files that draw the charts, so the words and the pictures cannot disagree.
Each chart names its source underneath and has a table with a download. Estimates are labelled where they appear. The full list, with retrieval dates, is in the Sources chapter.
Three things are kept apart on purpose: payments made in error, which agencies measure; fraud, which the Government Accountability Office estimates; and spending that works as the law says but that people argue about.
The model's current path
With no lever moved, the model should land where CBO's outlook does, in the debt and in the pieces that drive it. The first two tables compare the model with CBO's February 2026 projections, which were finished before the Supreme Court struck down the emergency tariffs. The third shows what CBO's later tariff estimate adds; the lab's current path includes it.
This tests the current path over thirty years. It does not yet test what the levers do over thirty years. CBO publishes long-term effects for five Social Security options and Social Security's actuaries publish 75-year effects for many more; checking the retirement age, wage cap and price-index levers against those is the next test to add.
Debt held by the public, percent of GDP
The pieces, thirty years out
Tariff changes since February
Each lever against CBO's score
The model was adjusted until it passed, as the change log below records, so the test shows only that the lab agrees with CBO over ten years. It does not show that either is right.
Our model's ten-year effect, as a share of CBO's
Like for like: CBO scored the same policy. Rough: CBO's policy differs from ours in a way the note explains. Scale only: shows what lever setting matches a CBO option, and tests nothing else. Rule of thumb: CBO's published sensitivity of the budget to growth or interest rates.
Show the numbers
Levers with nothing to check against
These have no comparable CBO score. Each is either plain arithmetic on a spending total or an assumption of ours, and is marked as such.
What the model does not do
It works in shares of the economy, one year at a time, with about forty assumptions. CBO's models track individual taxpayers and beneficiaries. So the lab is good for the size and direction of a choice and poor for detail: it cannot say who pays a tax, how a benefit change falls on different earners, or what happens in a recession.
The checks above cover the first ten years. Past that, the model extends the same relationships to 2056, and small differences compound. A point of extra growth held for thirty years is far outside anything CBO scores, and the result should be read as an illustration.
The site's public address is https://decisivelabsllc.com/atlas/.
Three outside reviews in November 2026 found the problems listed in the change log below. Every assumption is printed on the lab page.
Change log
November 15, 2026, version 0.18. Share captions.
- Every chart on the main page now has its own share caption, written from the data: a figure and its comparison, with no position on what should be done. The improper payments caption says that the total includes underpayments and that most of it is not fraud. No figure or assumption changed.
November 15, 2026, version 0.17. Sharing.
- Sharing a chart opens a dialog with an editable caption, a link, and a picture of the chart to put in a post. The caption states a finding from the data, such as how much of each dollar spent was borrowed, and takes no position on what should be done. No figure or assumption changed.
November 14, 2026, version 0.16. The lab's current path is now kept.
- A dated copy of the lab's path with every lever at its default (debt in 2027, 2030, 2036, 2046 and 2056) is saved and never rewritten; a new copy is added when the path changes or after 30 days. The main page lists each copy and marks each year as observed, with the published figure, or not yet observed. No assumption changed.
November 14, 2026, version 0.15. Additions after the same outside review. No lab assumption changed.
- The policy lab has a new section, "What it takes to hold the debt steady". It is budget arithmetic on the lab's current path with no economic feedback: debt next year is this year's debt times one plus the interest rate, divided by one plus growth, plus the deficit before interest. With every slider at zero it reproduces the lab's path exactly.
- Every chart on the main page and the lab can be shared as a link that keeps its settings, or saved as an image. The amount typed into the receipt is never put in a link.
- A "What changed" section lists figures that moved between daily refreshes, marking a new period apart from a revision.
- Every source line now says when its file was retrieved. On phones the section menu is one row that scrolls sideways; it was four rows.
November 14, 2026, version 0.14. Presentation changes on the main page after a further outside review. No figure or model assumption changed.
- The main page opens with the headline and the one-dollar bars; the opening paragraph moved below them.
- The receipt is now in two labelled parts: where your taxes went, which adds up to what you paid, and what was spent beyond taxes by borrowing, which is extra and not a bill.
- Charts drawn from the OMB Historical Tables now say which fiscal year is the latest and, separately, the day the file was retrieved.
November 14, 2026, version 0.13. More changes from the outside reviews.
- Tax changes phased in over three years while CBO's scores take effect at once, which made each point of tax worth about a tenth too much after the first decade. Tax levers now take effect in 2027 and were resized to CBO: a point of income tax raises 0.345 percent of GDP (was 0.37), a point of corporate tax 0.041 (was 0.045), a point of sales tax 0.17 (was 0.19), and the two wage-cap choices 0.211 and 0.415 (were 0.23 and 0.45).
- A carbon tax held its revenue as a fixed share of the economy for thirty years. It now loses 2 percent of that share each year as emissions fall. This is our assumption.
- Sponsor lists: the 2026 Social Security 2100 Act replaces the 2019 version, which is kept as a past proposal; the veterans entry now names the Take Care of America's Veterans Act and says it moves money within veterans programs; the sales tax is marked as scored with no sponsor; the tariff lever lists the Senate votes and the court ruling against the tariffs.
- Social Security and Medicare pages now give the trust fund depletion dates from the 2026 Trustees Reports.
- The fraud panel is retitled to say the estimate covers 2018 to 2022 and is not a current figure. The top of the home page now leads with fiscal 2026 totals from Treasury.
November 14, 2026, version 0.12. The site moved to decisivelabsllc.com/atlas/. No figure or assumption changed. The reviewers withdrew their finding that the SNAP figures were out of date: the site matches the Agriculture Department's October 2026 summary.
November 14, 2026, version 0.11. Changes made after three outside reviews (a budget analyst, a first-time reader and a sceptic).
- The current path was labelled "today's law" but followed CBO's February 2026 projections, finished before the Supreme Court ended the emergency tariffs. It now adds CBO's August 2026 estimate of tariff changes and says which projections it uses. Debt in 2056 on the current path rose from 173 to 179 percent of GDP.
- The plan "Grow out of it" set productivity a full point higher for thirty years, ten times the only change CBO has scored, and showed the result as a success. It is now "Assume faster growth", marked as not a policy, at half a point. The productivity lever runs from half a point slower to half a point faster. Any setting past 0.1 carries a notice and is not shown in green. Productivity was taken out of "A bit of everything".
- Extreme combinations could show debt below zero or in the thousands of percent, worded as wins. The lab now says such results are outside what the model can describe.
- The Medicaid lever acted on the whole health function, including research and public health agencies. It now acts on Medicaid, CHIP and premium credits only, which makes each point about a sixth smaller.
- The improper-payments lever treated measured errors as money that could be kept. It is now marked as an assumption, stops at 25 percent, and is out of every starting plan.
- The income-support lever included federal employee and military retirement. It no longer does.
- The plan "Raise revenue" was described as keeping benefits "as promised". It now says benefits are left as scheduled.
November 14, 2026, version 0.10. The current path is now also checked against CBO for revenue, spending, interest, Social Security and Medicare in 2056, and for the debt in six years. No assumption changed.
November 14, 2026, version 0.9. First test against CBO's scores, and the model was changed where it failed.
- Raising the retirement age saved about four times too much in the first decade. Savings now build slowly, because only new retirees are affected.
- The top wage-cap choice was "all wages" with no CBO score behind it. It is now CBO's option, taxing wages above $250,000 too, and both cap choices raise what CBO says they raise.
- A point on income tax rates raised about a fifth too little. Corrected.
- A carbon tax raised slightly too much. Corrected.
- Faster growth helped the budget less than half as much as CBO's rule of thumb. Extra growth now takes effect at once, brings in 22 cents of revenue per extra dollar of output, and benefits catch up with it over time. Interest rates now rise with faster growth, as CBO assumes.
- Rate changes reached the debt too slowly. Corrected against CBO's interest-rate rule of thumb.
These changes move the starting plans. "Grow out of it" now does better than before in the model, and plans that lean on the retirement age do worse in the early years.