Does the draw machine make a difference?

What is actually published

Every result in the operator's own export carries two extra columns. One identifies the machine, a physical drum. The other identifies the ball set, because a game does not use one set of balls but several, rotated between draws.

We hold 266 draws across four games with those columns filled in, drawn by 11 distinguishable machines. Nobody else in our archive publishes this. Spain, Japan, Brazil, Canada, Ireland and the United States all publish the numbers and nothing about the apparatus.

The identifiers are not consistent between games, which is itself a clue about who runs what. The UK's own games name their drums — Lotto1, Lotto4, Lotto5, Lotto6 for UK Lotto, the Excaliburs for Thunderball and Set For Life. EuroMillions uses bare numbers, because it is a nine-country game drawn in Paris rather than by the UK operator, and that draw identifies its machines and ball sets numerically.

Two things in that table are worth a sentence each. The first is that the UK games mix their drums and ball sets freely — UK Lotto has used 17 different combinations of 4 machines and 8 ball sets in the period we hold.

The second is that EuroMillions does not. Each of its 3 machines is loaded with the same ball set every time, in strict rotation. That is a fact about the draw, and it is also a limit on this analysis: for EuroMillions, "is this machine unusual" and "is this ball set unusual" are the same question, and the two tests below are one test reported twice rather than two pieces of evidence.

Some drums draw more than one game

The Excalibur machines draw both Thunderball and Set For Life. On either game's own page a drum looks lightly used; the drum itself has done rather more than that.

We count that, and we do not pool it. A ball numbered 40 in Thunderball is a different physical object from a ball numbered 40 in Set For Life, so combining the numbers themselves would be comparing two things that only look alike.

The test we cannot run, and why we say so first

The obvious approach is a frequency table per machine: count how often each number came out of each drum and test it against an even spread. It does not work at this sample size, and it is worth being precise about why rather than running it anyway.

A 6/59 game has 59 possible numbers and produces 6 of them per draw. For the expected count per number to reach five — the point below which the standard test is not merely weak but undefined — a single machine needs about 50 draws. For a table worth putting in front of a reader it needs about 197. The busiest UK Lotto drum has done 23.

So that table does not appear on this page. It is not a close call and it is not a judgement about presentation. On 23 draws the arithmetic produces a number, and the number means nothing.

The test we can run

Collapse 59 outcomes into two and the expected count per cell goes from about three to about ninety. That is a real test with real power, and it is not a consolation prize: it matches how a drum would actually fail. A ball that is fractionally heavier or larger than its neighbours sits differently in the drum, and the numbers printed on the balls run in order, so a physical fault shows up as a lean toward one end of the pool rather than as an affection for the number 17.

We use two collapses. High against low — for UK Lotto that is low is 1 to 29, high is 30 to 59 — catches a size or weight effect that tracks the printed number. Odd against even catches a grouping or paint effect that does not.

For each game we put the machines in the rows and the two buckets in the columns, and ask one question of the whole table: do these drums split the pool differently from one another? One question, one answer, per game per collapse.

What came back

What this does and does not rule out

It rules out a drum leaning noticeably toward one end of its pool, or toward odd or even numbers, at the sample sizes shown above. That is the failure mode a physically faulty drum has, and it is not there.

It does not rule out a drum with a small preference for one particular number, because nothing at this sample size could. And it does not, and never could, tell you anything about the next draw. Even a drum that genuinely favoured high numbers would not make a high number due, or a low one overdue. A machine has no memory and neither does a ball.

Method

Computed by scripts/machine-bias.mjs from the archive published on our data page. Draws are grouped by number matrix rather than by format record, so UK Lotto's June 2026 change — which added a second draw per night and left the 6/59 matrix alone — does not split a sample that is genuinely one experiment.

Pearson's chi-square on the machine-by-bucket table. A group holding too few balls to carry an expected count of 5 in both cells is excluded from the test and named below rather than quietly dropped, because a test that changed its own question without saying so is worse than one that refused to run.

• Thunderball, by ball set, high / low: excluded T8 (2 draws).
• Thunderball, by ball set, odd / even: excluded T8 (2 draws).
• Set For Life, by machine, high / low: excluded Excalibur1 (1 draw).
• Set For Life, by machine, odd / even: excluded Excalibur1 (1 draw).
• Set For Life, by ball set, high / low: excluded SFL1 (1 draw).
• Set For Life, by ball set, odd / even: excluded SFL1 (1 draw).

The threshold quoted in the verdict is Šidák's, which is what "the smallest of 16" has to beat before it means anything. Reporting the smallest p-value from a batch of tests without saying how many were run is the single most reliable way to find a pattern in a fair process, and we would rather not.

One correction we do make to the operator's own file: it contains both Lotto 6 and Lotto6, and one spelling of Excalibur 6. Runs of whitespace are collapsed so those read as one machine. Nothing else is normalised — names are not case-folded and no name is mapped onto another, because two spellings could be two drums.

Want the numbers instead of the apparatus? Generate a UK Lotto line, or read why hot and cold numbers do not work.

Last computed Sunday 20 September 2026 over 56 draws on the busiest drum (Excalibur5) and 266 in total.