Skip to main content
Story · exactness · 5 min
Month-end. The trial balance is off by €0.03. Not €30,000: three cents. Every individual transfer reconciles against the bank statement. Every journal, checked by hand, balances. The finance lead wants to know where three cents went, and the honest answer is that nobody has any idea.

​
Why three cents is worse than thirty thousand

A €30,000 break has a shape. It is one transfer, or one duplicated batch, and you find it by sorting on amount. Three cents has no shape. It could be one error of three cents, three errors of one cent, or three hundred errors of one hundredth of a cent that happened to accumulate. There is nothing to sort on, nothing to filter by, and no transaction that looks wrong.
The dangerous property of a small break is not its size. It is that it grows with volume: and that it is indistinguishable, on any given day, from a system that is working.
Someone will eventually suggest a tolerance. That suggestion is the real failure, and it arrives about a week in.

​
The diagnosis

The fee schedule is 1.5% on transfers below €50. Someone sends €33.33.
Not representable in cents. Something has to happen. The code does the natural thing:
€0.49 is booked as revenue, the customer is debited €33.33, and €32.84 moves into in-transit. The journal is written and the ledger accepts it, because €33.33 = €0.49 + €32.84 exactly. And €0.00995 has quietly ceased to exist. Not moved. Not lost to a rounding account. It was in the multiplication result and it is not in the journal, and no line in the system records the discrepancy, because from the journal’s point of view there is no discrepancy. The arithmetic closes. It just closes around a different amount than the fee schedule specifies. Three hundred transfers of that shape in a month is €2.98. Round it the other way on some and the residuals partly cancel, which is worse: the break becomes small, erratic, and impossible to characterise.

​
Where it actually goes

The residual does not disappear into the void; it silently changes who owns it. Round the fee down and the customer keeps a fraction they were not entitled to. Round it up and Arc takes a fraction it did not charge for. Over a month, across a mixed book, the direction depends on the distribution of transfer amounts. Nobody decided this. It is a policy the system is executing, that nobody wrote, and that changes with traffic.

​
The fix, and why it is not “round better”

There is no rounding mode that makes 0.49995 fit in cents. The residual is real. The only question is where it is recorded. Arc records it:
divRound returns the rounded value under an explicitly chosen mode. divResidual returns the exact leftover, as a bigint. It exists for precisely one purpose: so the leftover can be given somewhere to go.
The journal still balances. The fee is still €0.49. But the residual is now a number someone can look at: a balance on revenue.rounding.EUR that can be queried, reported, and explained.Three cents at month-end stops being a mystery and becomes a line item.

​
What made it possible

The residual mechanism only works because of a decision made much earlier: money is a bigint count of minor units, never a float.
With floatsamount * 0.015 produces a value that is already slightly wrong before any rounding happens. There is no exact residual to extract, because there was no exact result.The best you can do is measure the discrepancy against another approximation.
With integers3333n × 150n / 10000n has an exact quotient and an exact remainder. divResidual returns the remainder because the remainder genuinely exists.Exactness is what makes the residual nameable.
This is why the lint rules are aggressive. Math.round, parseFloat, toFixed and fractional literals are all build errors in source, not because any single use is catastrophic, but because each one silently removes the ability to account for what it discarded.

​
The property that pins it

One example does not prove an allocator is safe. So the guarantee is stated as a property and tested against generated inputs:
Splitting any amount across any weights always sums back to exactly the original.
The naive implementation fails this immediately. Splitting €10.00 three ways gives €3.33 each, totalling €9.99: a cent evaporates. Arc’s allocator distributes remainders deterministically so the total is preserved, and the property test asserts it across randomly generated amounts and weight vectors rather than the handful of cases someone thought to write down.

​
The lesson

1

Rounding is a policy decision, not an implementation detail

Every rounding operation decides who keeps the fraction. Making the mode explicit forces someone to choose deliberately rather than inherit whatever the language does.
2

A residual must have somewhere to go

“Round and move on” is not an algorithm: it is an unrecorded transfer of value. Give the remainder an account and it becomes auditable.
3

Never accept a tolerance on the balance check

The moment Math.abs(debits - credits) < 0.01 appears, the ledger stops being a ledger. There is no defensible answer to “how far off is acceptable?”, and the threshold only ever moves in one direction.

Why money is never a float

The mechanism, specified rather than narrated.

The journal that balanced and lied

Next: a ledger where the arithmetic was perfect and the record was false.