Round numbers are the tell: how to spot an invented rate table
While auditing 102 tax models we needed a way to tell a real schedule from an invented one at a glance. The most reliable signal turned out to be aesthetic: real tax law does not produce round numbers.
· Valem
The signal
Denmark's vehicle tax model computed its charge as: 400 if the car does at least 20 km per litre, 800 at 15, 1,600 at 10, and 2,600 otherwise. Four round numbers, doubling neatly, standing in for a thirty-band official table.
Nothing about that expression is broken. It is monotonic, it is roughly the right order of magnitude, and it passes every test written against it. But no legislature has ever produced a schedule that looks like that. Statutory tables are the residue of indexation, negotiation and rounding conventions applied over decades: they contain values like €57.60, 146,901 CZK, 2,818,152 DKK and 68,904. Cleanliness is evidence of authorship.
What we did about it
We classified every vehicle-tax model as either asserting a statutory schedule or reproducing only the shape of one, and made the second group say so. The disclosure is a warning at the top of the model, in the page's own language, lifted onto the static page so it is visible without JavaScript; the badge changes from "verified" to "illustrative rates"; and the disclaimer states plainly that the figure will not match a real bill.
Sourcing the real schedule where one existed changed the amounts substantially, which is the point. Luxembourg's statutory formula is a simple product with a €30 floor, and where the invented table had charged €140 the law gives €57.60 — more than double. Portugal, Sweden and Singapore each published a worked example, and all three of our conversions reproduced them to the cent, which is what makes a rewrite trustworthy rather than another guess.
The classifier that over-matched
Our first attempt at finding these keyed on the word "estimate" in each page's own language, and it produced false positives worth recording.
Japan's note used 概算, which does mean estimate — but the model had the real statutory displacement table all along. Austria's figures were a genuine formula with drifted constants, where the annual deductions fall each year from 65/115 to 62/106. Singapore and Slovenia both had fully published formulas we could simply implement. And Croatia's case was the reverse of what the classifier suggested: the rates were fine and the inputs were wrong, because the fee is charged by engine displacement rather than by power and age.
A word in a note is a weak signal. The shape of the numbers is a strong one.
A second signal: the table that is too short
Round numbers travel with a companion tell, which is a table with too few rows. Real vehicle taxes are charged in bands because bands are how legislatures compromise, and the band count is rarely tidy: Denmark publishes thirty, Japan charges by engine displacement across a full statutory ladder, and Portugal combines a displacement table with a CO2 table, multiplies by a year coefficient and then adds a diesel surcharge.
A four-row table standing in for any of those is not a simplification of the law, it is a different law. The distinction matters because a simplification degrades gracefully — it is roughly right in the middle of the range and drifts at the edges — whereas a substituted table can be wrong by a multiple anywhere, with no pattern a reader could correct for.
The same reasoning applies to the inputs. If a model asks for one number where the statute reads three, the missing two have been assumed, and an assumption inside a rate table is indistinguishable from a fact once it is rendered as a figure with a currency symbol on it.
Where an honest "we do not know" is the right answer
Eighteen of our 102 pages still carry the illustrative disclosure, and they are hard for structural rather than effort reasons. Switzerland has twenty-six cantonal systems and Belgium three regional ones; Bulgaria, South Africa and Australia set the charge municipally, provincially and by state. There is no national schedule to ship, so any single number a page displayed would be a fiction.
The rest exist but are not reachable: Finland's sits in a government decree, Malta's and Turkey's official tables return 403 to us, and Denmark's thirty-band schedule is only exposed through the vehicle register's own lookup. Each disclosure names the exact document that would unblock it, so the next attempt starts from an address rather than a search.
We would rather ship a page that says what it does not know than one that quietly rounds it off to something plausible. Plausible is the failure mode.
Related
We audited 102 tax calculators. 29 were wrong.
Three ways a calculator is wrong that its own tests can never catch
See the 102 calculators this came from · All research
Built with Valem — models you can read, re-run and check.