The two checks, and what neither one proves
The cheapest thing on this site for a skeptic to verify, so it is worth stating precisely rather than impressively. There are two checks, they are different animals, and we used to describe them as one.
- Completeness
- 114 of 114 months · worst deviation 0.000%
- What it compares
- The sum of our per-country Census rows against the same Census API’s own world-total row, same month, same vintage.
- What it proves
- The pull is complete and the country filter is right — on every month from 2017-01 to 2026-06.
- What it does not prove
- Anything external. It cannot fail unless the pipeline dropped rows. It is not a reconciliation, and Treasury is not in it.
- Agreement with Treasury
- A different check with a different number: duty assessed at entry ÷ Treasury gross receipts sits at 0.85–0.95 in every year, 2017 to 2026.
The assessed book excludes anti-dumping and countervailing duties and fees, and entry month is not cash month. The gap is explained, not residual.
Why gross, not net
From May 2026 court-ordered IEEPA refunds drove Treasury net toward zero, so the same ratio reads 1.78 on the net book for 2026. That is a vanishing denominator, not a finding.
Agreement is not reconciliation
No Treasury figure is used to build these numbers. Two books that were never joined landing inside an explained band is worth something — not the word reconciled.
Until the 2026-08-26 audit the ledger said: “Reconciled against US Treasury monthly customs receipts (yearly ratio 0.89–1.03) and Census world-total rows (0.000% deviation, 114/114 months).”
Two errors in one sentence. Calling the 0.000% a reconciliation invited a reader to believe an external check had closed to zero; none had. And 0.89–1.03 was not reproducible by any statistic — it mixed a ratio-of-sums against the net-of-refunds book with a mean of monthly ratios. Against gross the honest band is 0.853–0.947 in every year including 2026. Logged, and one machine-readable surface still serves the old string — logged too, rather than quietly patched.
Check it yourself
round(sum(b.duty)/sum(m.gross_usd),3) AS ratio_of_sums,
round(avg(b.duty/m.gross_usd),3) AS mean_of_monthly
FROM burden_national b JOIN mts_customs m USING(month) GROUP BY yr;
Both inputs are public and neither is ours. The assessed side is our CSV, or re-pull it from Census; the Treasury side is MTS Table 4, classification Customs Duties, no key. If a year lands outside 0.85–0.95, one of us is wrong and we would like to know which.
One limitation, and it is ours
The Census world-total rows are not stored, so the completeness check is not re-runnable from our files — you would have to re-pull those rows to audit it, and until we store them you are taking 0.000% partly on trust. The 2026-08-26 audit recommended storing them. It has not been done.
Two months have been independently re-pulled and matched to the dollar, 2019-06 and 2023-03. Two of 114 is a spot check, not a verification, and is described here as one.
how the numbers are made · what we got wrong · what we tested and it failed · check one of your own entries