What the record shows

THE LEAD · 2024-01 → 2026-05 · all origins · US customs records, monthly

The trade war moved exactly one number

The tariff rate at the border doubled, then gave a third back — and the structure of US trade barely noticed. One share fell, four absorbed it, and everything else held to the decimal.

The whole measurable structural effect of the trade war is one number: China's share of containerized import value fell 29.9% → 19.4%, and that share flowed to Vietnam, Thailand, Indonesia and Cambodia. Between 2024 and mid-2026 the average tariff rate collected at the border roughly doubled, then gave a third of it back — and almost nothing else in the structure of US trade moved: the country-by-commodity pattern of imports correlates at 0.826 with itself month to month, and 0.772 across the entire war window (1.0 would mean the map of who-ships-what never changed at all). The rest of the mix did not diversify — outside China the supplier mix was flat, 4.88 → 4.86 bits on a standard scale of how evenly trade spreads across suppliers — and the total pie shrank 12.6% in declared value. Leaving China had a measured price: the $61.2bn of trade that moved was bought about 17.3% higher up each good's price range than the Chinese dollars it replaced — one substitution, priced, on a structure otherwise too rigid for the largest tariff shock in a century to bend.

Shares and correlations from monthly US customs import records by origin and six-digit product code — the customs system's ID for a kind of good; the collected-rate arc from official duty receipts. Recomputed daily.

What would prove it wrong: China's share regaining half its loss, or the ex-China supplier mix widening materially — either breaks “one move.” The tempting “diversification” reading was tested and killed: China's decline alone explains the entire shift in the supplier mix.

→ the check prices each of your lines at its own date and origin — /check

02 · published 2026-07-28 · confirmed 2026-08-29 · two independent ledgers

We found an error in IMF data. They confirmed it in writing

Confirmed in writing · IMF PortWatch · 2026-08-29

The IMF's PortWatch platform showed vessel calls at India's main container ports collapsing 40–52% from March 2025. US customs records showed imports from India flat to strongly up across the same window — and goods cannot reach the US without leaving India. We published the disagreement with the test that would prove us wrong, then reported it directly to the PortWatch team. On 2026-08-29 they replied: “We looked into the issue and found that it stems from irregularities in the underlying AIS data source, over which we have no control” — and added a note on the anomaly to their public FAQ. The method that caught it is the same one that scores your entries: two independent records of the same reality, compared, with the disagreement published rather than smoothed over.

Blind comparison of the customs import record against the vessel-call series, indexed to a common start; archived copies of the published values by retrieval date, available on request.

What would prove it wrong: Evidence that Indian container exports to destinations other than the United States fell by roughly 40% in this window while US-bound trade held up. That would make both instruments right. It requires a third source — EU or intra-Asia import statistics — which we do not yet hold.

Confirmed in writing by the IMF PortWatch team, 2026-08-29; a note on the anomaly was added to their public FAQ. The full correspondence: /findings/f-001.

→ the same two-ledger comparison runs on every entry sent — /check

03 · 2026-06 · China · six test goods · duty receipts vs the written schedule

The schedule cannot be added up

A Chinese good can carry several tariff measures at once — stickers on the same box, each with its own rate. The four stickers that sum to 75% on paper collected 46.0% on sweaters and 10.0% on toys. The ratio of collected to schedule-sum runs from 0.64 down to 0.05 — an 11.9× spread, where a fixable arithmetic error would be a constant. Exclusions, effective dates and scope annexes are the whole difference, and none of them live in the addable part of the schedule. This is why every published “effective tariff rate” disagrees with every other — and why this site publishes what was actually collected.

What would prove it wrong: a single scalar mapping schedule sums onto collected rates within ±5 points on all six codes. Registered before the run; the best case missed by more than 25 points.

→ the check composes the rulebook per line instead of summing stickers — /check

04 · 2024-01 → 2026-06 · ten origins · 1.02M priced lines

How faithfully does the border collect its own rulebook?

The rate collected at the border runs below the rate the rulebook required — by a small, persistent, mapped amount, part of it lawful. We recomputed the required duty on 1.02 million lines — each one product code, from one origin, in one month — and compared it with what the receipts show was collected. The gap is not noise: a line that misses in one month misses the same way in months held out of the fit — the sizes of the misses match at correlation 0.978 — and where the gap is meaningful, its direction repeats 98.5% of the time, where always guessing the commoner direction would score 81.9%. And the gap grew with the rulebook: as the required rate roughly doubled (11.6 → 25.3 percentage points), the relative gap also doubled — 7.2% → 14.1% of the required rate — and 92.9% of the duty on lines present in both eras got worse. Complexity degrades collection. Part of the gap is lawful (exclusions filed at entry, duties charged on only part of a product's value — its steel content, say — and filing timing) and the public record cannot fully separate lawful from error — which is exactly why the honest unit of output is a per-line disagreement with its reason, not a total.

|required − collected| as a share of required, duty-weighted (lines carrying more duty count for more), on determinable lines only. Eras: 2024 vs 2025-05→2026-06. Recomputed daily.

What would prove it wrong: lines losing their direction between the months used to build the map and the months held back to test it — the map decaying toward chance — or the relative gap failing to move with rulebook complexity. A pre-registered test of whether this gap predicts recoverable money failed, and is published.

→ this map is drawn nation-wide; the check draws it for your book, line by line — /check

05 · 2024 vs 2026 · 1228 goods tested · declared value per kg, by origin

What should a thing cost? The record already knows

Take any traded good and look at the price per kilogram declared by every country that ships it. That spread is not random — it is a signature of the good itself. Across 1228 goods with at least fifteen origin prices in both years, the 2024 spread predicts the 2026 spread at correlation 0.776 — and the most any signature could score here is 0.937, the value you get comparing one year's data against itself, so this is about as stable as the data can show. Knowing only what the good is explains 56.2% of all price variation across lines, and 84.3% where the trade value is: where the money is, the commodity tells you the price.

Cross-origin spread of log declared value per kg, per code per month (months with ≥15 origins), averaged by year. Stability = correlation of the per-good spread across years. Recomputed daily.

What would prove it wrong: the 2024 spreads failing to predict the 2026 spreads. The correlation above is re-measured on every refresh and prints whatever it is.

→ this yardstick already scores every line of an uploaded entry — /check

06 · 2024-04 → 2026-04 · matched goods · 15.7M customs records

The shrinking Chinese invoice

Across 136 matched commodity codes, the declared customs value of Chinese cargo fell 23% per kilogram between the year before the April 2025 tariff wave and the year after — the same goods, a quarter less value on paper — against −9.5% for India and −7.5% for Vietnam. It replicates on three cuts of the data not used to find it, and the gap on identical goods is now extreme: wooden furniture leaves China at $0.65/kg and Vietnam at $2.36/kg under the same code. The mechanism is deliberately not claimed. Four candidates — discounting, undervaluation, product downgrading, origin-washing — remain open; the strongest alternative (Chinese cargo relabeled through Vietnam) was tested the day it was raised and covers under 5% of the decline. The fall is national-scale: the containerized import bill fell about $102 billion in a year on only −3.5% of weight — and split the fall into volume and price, China accounts for effectively the entire net decline, 70% of it price rather than volume.

Value ÷ weight within matched six-digit codes above 200 kt, year before vs year after 2025-04, from US customs records. Recomputed daily.

What would prove it wrong: a matched-code decline of similar size in India or Vietnam (not observed), or the Vietnam-rerouting channel explaining the majority (tested: under 5%). Registered with two kill-switches before the run; both held.

→ declared value is exactly what the check scores, per line — /check

07 · 2024-07 → 2026-06 · one industry · duty receipts, code 847130

A tariff moved laptops out of China. Repeal brought 7% back

China's share of US laptop imports went 65% (February 2025) → 7% (June 2025) under an 18.8% collected burden — bottoming near 2% that autumn — while world totals held within ±18%: every dollar visible arriving instead from Vietnam, Taiwan and Mexico. Then the tariff came off. By June 2026 China's share had climbed back only to 11% — of the 58 points of share it lost, it regained 4: 7% of the loss. Supply chains are a one-way door. And the industry left China for Vietnam, Taiwan and Mexico — not for the United States. Nothing here shows reshoring.

Origin shares of monthly import value on code 847130; the burden is duty ÷ dutiable value from official receipts. Recomputed daily.

What would prove it wrong: recovery: the registered bar was China regaining 50% of the lost share within a year of repeal. Measured: 7%.

→ every rate in this story is a collected rate; the check reads the required one — /check

08 · 2024-01 → 2026-06 · eleven chapters · duty receipts by chapter

When the retreat came, one statute melted and the other did not

The 2026 de-escalation did not lower tariffs across the board — it sorted by legal authority — by which law each tariff was written under. Sectors under Section 232 kept their increases: aluminium 100% retained, steel articles 88%, autos 72%. The IEEPA-driven chapters gave half or more back: machinery 47%, electronics 45%, toys 28% — furniture fell below its 2024 baseline. You can read the statute from the cash register, and which authority your goods sit under is the single best predictor of whether their tariff survives the next court ruling.

Chapter-level collected burden: 2024 baseline vs the 2025-09→2026-01 peak vs 2026-04→06, retention = share of the increase still standing. Recomputed daily.

What would prove it wrong: any overlap between the two groups' retention ranges. Measured: 70–100% for the Section 232 sectors against -9% to 47% for the rest.

→ the check names each line's legal authority from its overlay codes — /check