65,637 returns, read as 11,842 chains.

Each year is re-done under three rules, none ours. A year’s qualifying distributions go first to the prior year’s undistributed income, then to the year’s own amount, then to corpus (IRC §4942(h)). The carryover applied is the lesser of the excess carried in and what the year’s own distributions left, the earliest excess first (Treas. Reg. §53.4942(a)-3(e)(1)). What is left of an excess after its fifth year is gone ((e)(3)). An election the return states on lines 4b, 4c or 7 is taken as filed.

A return opens where the last one closed: last year’s undistributed income is this year’s line 2a, and last year’s carryover, year by year, is this year’s line 3. 50,339 pairs of consecutive returns were held against each other.

Where a carryover opens above the prior closing, $1.3B appears across 527 links; where it opens below, $878M is dropped across 501. A return may rest on something these files do not show: an amended return not yet published, a corrected valuation, a year filed on paper. 2,483 pairs were not consecutive periods and were not compared: a gap is said, never bridged.

6,093 of 10,947 foundations carry an excess out of their latest return. The median holds 1.15 years of its own distributable amount. An excess lasts five years. 3,274 foundations hold $21.2B in its last: applied in the next tax year, or gone.

A lapse is not a shortfall. It is a credit that ran out. It matters to the foundation that will one day distribute less than its amount and would have had the excess to set against it. Each year far more excess arises than is ever applied, so the amount carried grows and the oldest part of it runs out. Years with fewer than 5,000 returns read are left out; the latest year is still being filed, and few returns for years before 2020 were filed electronically.

Size is total assets at fair market value on the latest return. “Years of payout” is the excess carried, divided by the latest year’s distributable amount: among foundations that carry one, a tenth hold under 0.08 years of it and a tenth hold more than 13.5.

• 273 filed years: carryover applied ahead of the year's own distributions, or beyond what was left undistributed.
• 173 filed years: the lapse on line 8 is not the unused excess from five years back.
• 131 filed years: the year's distributions applied in another order.
• 130 filed years: lines 4a to 4e do not add up to line 4.
• 108 filed years: carryover left unapplied while the year's amount stayed undistributed.
• 104 filed years: the years on line 10 do not add up to line 9.
• 82 filed years: one of the form's own sums on lines 6 and 9.
• 32 filed years: the prior year's undistributed income not reduced first.
• 6 filed years: line 3f is not the sum of lines 3a to 3e.

A year may show more than one. The first is the one the form’s instructions caution against by name: an old excess applied to the year’s amount while the year’s own distributions are elected to corpus comes out as a new excess with five fresh years. Each is the return’s figure beside the recomputed one, as filed and as recomputed; the return may carry a reason the file does not show, such as an election statement. The pre-filing check re-adds the same lines on a draft, in the preparer’s own browser.

The IRS’s Statistics of Income Division tabulates the same returns each year, and includes every return with $10 million or more in assets. For tax year 2022, its Table 1 and the e-files, totalled the same way, for foundations of $10 million and over:

The IRS’s table: Table 1. Domestic Private Foundations: Number and Selected Financial Data, by Type of Foundation and Size of Fair Market Value of Total Assets, Tax Year 2022 (IRS, Statistics of Income Division, Domestic Private Foundations Study, February 2026.) irs.gov, SHA-256 030ed8abaed6185f…. The e-files: 122,932 Forms 990-PF for tax periods ending 2022-12 to 2023-11, one return a foundation; set aside as the IRS sets them aside: 94 foreign organizations, 5,829 nonexempt charitable trusts (its Table 2), 1,085 earlier or duplicate submissions, and 1,001 published after its two-year window closed. A further 4,258 returns report nothing at all (no assets at either end of the year, no revenue, no distributions) and are counted apart; the IRS does not say how it treats them, and with them left out the two counts meet: 110,456 here, 109,351 in the table. SOI's note [2]: one foundation's total assets are reduced by $67.3 billion, its interest in its own trust, which files its own return. The same reduction is applied to the same return here.

Under $10 million the IRS samples, so its figures there are estimates and a difference is expected. It is widest in qualifying distributions: the sample is drawn by assets, and a foundation that receives a large sum and distributes it within the year ends the year holding little. From $10 million up the IRS counts every return, and the two agree to within a fraction of a percent on every line but the carryover. What is left fits what the IRS says of its own table: its data are “subjected to comprehensive testing and correction procedures”, and changes a filer makes by amendment “are captured, if available”.

• The totals, as one file: data.json. As tables: by tax year, by size, Part XII line by line, the reconciliation with the IRS’s table. CC BY 4.0.
• Any foundation’s returns, from irs.gov, with nothing from us: irs990pf_fetch.py reads the IRS’s yearly indexes, finds each return inside the published zips and fetches that one file. Standard library only.
• Any folder of returns, read as a chain: verify_chain.py re-adds every Part XII line and holds each year against the one before. It shares no code with this site; the figures here were held to it on real returns. It also checks a sealed chain: the body’s hash, each return’s bytes, and every year re-derived.
• The rules are the statute’s, the regulation’s and the form instructions’, each named above; both programs state them at the top of the file. No per-foundation table is published: a foundation’s page exists when someone asks for it, and is not listed.

python3 irs990pf_fetch.py 12-3456789 --out returns/ && python3 verify_chain.py --read returns/

• The population: every private foundation in the directory (11,842 with at least $10M in assets, as of 2026-09-26) and every Form 990-PF the IRS has published for it as e-file data: 65,637 returns, tax years ending 2020 to 2025 for the most part. Where a period was filed more than once the latest submission stands (973 earlier ones replaced).
• Each return read from the IRS’s own file, its bytes checked, and re-done with exact decimal arithmetic. Nothing is estimated and nothing is sampled.
• The arithmetic is held to the regulation’s own worked example and the form instructions’ examples, and to fifteen real returns read a second time by an independent program that shares no code with the first.
• Any foundation’s chain can be reproduced: seal its return, and the proof page reads its filed years in order and seals them, on request, under one hash.
• Totals only. No foundation is named here, and nothing here counts a late distribution; the form’s own taxable-amount lines are re-added like any other line and never tallied.
• The totals are published as data under CC BY 4.0. The IRS data is public domain. Cite as: ParetoAlpha, “The Carryover Chain,” 2026-09-30.