◀ THE FOLD0ROOT.AI // WORLD II · LOOT · THE MINT◆ .dlw.fold
THE FOLD / LOOT / THE MINT / THE ROUND HALF EVEN

THE ROUND HALF EVEN

exactly half a unit, every time
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Exactly half. The rule everyone learned is to round up, and applied to a column of money it quietly adds a tenth of a penny to every tie, in the same direction, forever.

LIT verified live. rounding the 1,000 exact ties 0.5, 1.5, 2.5 … 999.5, whose true sum is 500,000: rounding half up gives 500,500, a bias of +500 — exactly 0.5 per tie, every time, in one direction. Rounding half to even gives 500,000, a bias of 0. Half-up sends 0.5 to 1, 1.5 to 2 and 2.5 to 3; half-even sends 2.5 to 2. And Math.round(−0.5) returns −0, which is a different value from 0 to five of the operations that could receive it.
2 HOW IT WAS WEAVED · AI + HUMAN
Round half to even is the IEEE 754 default and is called banker’s rounding for the obvious reason; half-up is what school taught and what most naive implementations do.

AVAN (AI) summed a thousand ties rather than arguing about fairness, because the bias is not statistical — it is exactly 0.5 per tie with no variance at all. That matters: a random error averages out over a long ledger and this one accumulates linearly, so the discrepancy grows with the size of the business rather than shrinking with it.
3 ONE DIMENSION
A thousand ties. Half-up is off by exactly 500.
4 TWO DIMENSIONS · INTERACTIVE
Add ties to the column and watch the two totals separate.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object.
AVAN’s addition (the inverse-companion): the forward reading is that banker’s rounding removes the bias. The inverse is that there is no unbiased way to round a single number — the fairness is a property of the column, not the value. Half-even is not more accurate about 2.5; it is wrong about it by the same half. What it does is arrange for the errors to point in opposite directions often enough to cancel. Read backwards, this is a rule whose whole justification only exists in aggregate, applied one value at a time by code that can never see the aggregate.
LIT rounding the 1,000 exact ties 0.5, 1.5, 2.5 through 999.5 whose true sum is 500,000: rounding half up gives 500,500, a bias of +500 - exactly 0.5 per tie, every time, in one direction - while rounding half to even gives 500,000, a bias of 0; half-up sends 0.5 to 1, 1.5 to 2 and 2.5 to 3 where half-even sends 2.5 to 2, and Math.round(-0.5) returns -0, a different value from 0 to five of the operations that could receive it

FIG Round half to even is the IEEE 754 default and is called banker's rounding for the obvious reason; half-up is what school taught and what most naive implementations do. AVAN summed a thousand ties rather than arguing about fairness, because the bias is not statistical - it is exactly 0.5 per tie with no variance at all. That matters: a random error averages out over a long ledger and this one accumulates linearly, so the discrepancy grows with the size of the business rather than shrinking with it.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE MINT · David Lee Wise (ROOT0), with AVAN