What we tested, and it failed
A party with a fee riding on the finding does not publish this page. That is the only reason it is worth your time.
2026-08-28 · the rule gap does not predict recoverable money
- The registered question
- Does our own headline signal — the gap between what the composed rulebook required and what was paid — find the codes where duty was actually struck down?
- Answer
- No. ρ +0.251, p = 0.178
- Beaten by
- Sorting the same codes by how much duty they paid: ρ +0.356. No engine required.
- Where the money lands
- Top ten codes by our signal hold 36% of the struck dollars. Top ten by duty paid hold 65%.
- The book
- 86 HS6 codes from a 436-entry book, India, $777,512 struck across 36 of them.
This is the test the desk’s own contract named, in these words, as its proof of value.
The largest code
The single biggest struck code ranks 22nd of the 64 codes our all-cells signal reaches. Four of the top twelve carry no signal from it at all.
| ranking the book’s codes by | codes ranked | ρ | p | top-10 $ captured |
|---|---|---|---|---|
| duty paid — a sort, not a model | 86 of 86 | +0.356 | — | 65% |
| our rule gap, ruling-resolved cells | 29 of 86 | +0.251 | 0.178 | 36% |
| our rule gap, all cells | 64 of 86 | +0.160 | 0.201 | 2% |
| the ledger’s deviation | 85 of 86 | +0.178 | 0.100 | 15% |
These are four different populations, not four rankings of one. Our signal only exists where it can be resolved, so the row it is judged on covers 29 of the 86 codes while the sort it loses to covers all 86. That asymmetry is part of the finding rather than a footnote to it: a signal that is silent on two-thirds of the book has already lost to one that is never silent. It also means the p-values are not comparable across rows, and that ρ = +0.251 on 29 codes is a weaker statement than the same number on 86 would be — at 86 it would read p ≈ 0.02 and would not be a null at all.
The caveat cuts our way and travels anyway: struck means duty paid under a measure a court later invalidated — India’s 50% penalty, struck 2026-02-20. A correct filing carries it, so this is a legal event rather than a filing error, and a rate-variance signal was arguably never the right instrument for it. Decisive regardless: it is the gate we ourselves named, and we lost to a sort anyone can do in a spreadsheet.
What it settles: we will never tell you the gap predicts what you can recover, and nothing here is framed as refund-finding. What it leaves open: whether the gap measures something else real. It is independent machinery — the two signals disagree by more than 5pp on 81% of 144,761 shared cells. 5pp is a threshold we chose, not a fact, and the 81% moves with it; read it as "these two instruments mostly do not agree", which is the claim being made, rather than as a measured constant. Just not machinery pointed at money.
The Red Sea crisis did not move cargo between US coasts
Registered: Suez avoidance adds 10–14 days to Asia–East Coast transits, so East Coast share should fall. It ran 56–60% through the disruption window (56–61% across the whole plotted period), with pre- and post-window means both at 58.0%. The spread is 5.7 points over the wider window against a 5-point band, so the verdict turns partly on rounding — and month-to-month scatter of about 2.8 points means this design can only exclude coast shifts larger than roughly ±2 points of share. Nothing bigger than that happened; something smaller could have.
The 24/7 ports push left no trace in the physics
Registered with the direction stated in advance: night-hour berthing at LA–Long Beach should rise at least 4 points more than at untreated gateways. It went 36.0% → 34.0% against a control flat at 30.1% — −2.0 against a bar of +4, n = 21,700 calls. The study also preserves the mistake: an exploratory probe first showed a dramatic +9, which was UTC hours read as local time.
The “dodge premium” mostly is not real
Registered: trade migrates to the lower-tariff origin in proportion to the rate spread. Vietnam grew fastest where the spread was smallest; the monotone relation the claim requires is absent. The lesson outlived it — rate-on-dutiable is computed on a rounding error wherever the duty-free share is high, manufacturing spreads on codes where almost nothing pays duty. The study, with its own defect printed rather than left to be found: the served bucket table has no script on disk and cannot currently be reproduced.
The claim this page is most vulnerable on
“Registered before the data was touched” is what makes a null worth anything, and our own 2026-08-26 audit found it not supported by the file record for at least 12 of 20 studies — results dated before the registration block, falsifier fields containing outcomes, thresholds set knowing the result. One is cleanly pre-registered; the rest are honest analyses carrying a claim they cannot fully support.
The studies stand — their arithmetic was separately recomputed. What is withdrawn is the strength of the claim, which is what a correction is. A hostile reader would have found this anyway; better on our own page.
what we got wrong · the conflict, stated first · the findings that did hold · check one of your own entries