Skip to content

The dice are fair. Here's the proof.

Every craps app gets accused of rigging the dice, and most answer with a shrug. We can answer it twice: the engine that throws the dice has no access to your bets, and we measured 10,000 real throws to confirm the distribution is what a fair pair produces.

Download the full report (PDF)

Proof one: the design

The dice cannot see your bets

The usual accusation is that the game watches what you wager and adjusts. In Press It! that is not a policy we promise to follow. It is something the code cannot do.

Rolls come from a physics simulation: two rigid bodies thrown into a walled table with the velocity, spin and angle varied the way a human wrist varies them. Bullet drives it on Android, SceneKit on iOS. Whatever face is up when they stop rolling is your number. Nothing selects a total in advance and animates dice to agree with it.

The physics engine is a sealed module. Your bankroll, your bets and the state of the table are simply not passed into it. There is no wiring through which the throw could learn what you have riding on it, even if it wanted to.

The statistics are held to the same standard. Only genuine physics throws are counted; scripted tutorial rolls and debug rolls are excluded from the tally, so the dashboard describes the real engine and nothing else.

No two throws start the same

A physics table that launched the dice identically every time would be a machine for repeating itself, however good the simulation. So the throw is varied the way a human hand varies it, and four things are drawn fresh for every roll. All of them seed the launch. None of them touch the result, which is read off the dice after they stop.

Hand and wrist position
The launch point is jittered up to 0.75 m either side of centre, and the two dice leave the hand 0.56 m apart, so they never start stacked the same way twice.
Throw velocity
A forward impulse of roughly 100 units toward the far wall, give or take 7.5, on an upward arc of 4 to 6 with about a unit of lateral drift. The second die gets its own independent offset on all three axes, so the pair never flies as a unit.
Wrist snap
An angular impulse of up to 7.5 on each of the three axes, drawn independently for each die. This is the spin, and it is what makes them tumble rather than slide.
Starting orientation
Spun to a random yaw about the vertical. Left to itself the pitch and roll are random too; set the dice yourself and the face you chose starts up instead.

And what does not vary

Gravity
18 m/s²
Die
0.8 m, mass 2.5
Bounce
0.55 off a die, 0.45 off a wall
Friction
0.65 on a die, 0.6 on a wall
Damping
0.3 linear, 0.5 angular
Table
20 m × 6.5 m, walls 10 m high
Settled
both dice under 0.4 m/s for 25 straight frames
Result
whichever face points up at rest

The same table, the same cloth, the same gravity for everyone. Those figures are matched between the Bullet engine on Android and SceneKit on iOS so a roll behaves the same on either phone.

Set the dice, like you would at a real table

Tap either die in the heads-up display and it cycles through one to six. Whatever you leave showing is the face that starts up when you shoot, so you can set the 3V, the hardways, or whatever grip you swear by, and watch it leave your hand that way.

This is a real initial condition, not a decal. The die body is genuinely rotated to your chosen face before the impulse is applied, exactly as if you had placed it that way. Then it is thrown, it tumbles the length of the table, it hits a wall, and the number you get is whichever face happens to be pointing up when it stops.

What the published figures cover

Worth being exact about. The 10,000-roll validation below was thrown with the starting orientation randomised, which is how the dice behave when you do not set them yourself. That is the configuration those figures describe.

Proof two: the measurement

10,000 rolls, chi-square tested

A physics table could still be biased by its own geometry, so we did not take the design on faith either. An instrumented test drives the production engine headlessly, throws the dice 10,000 times and records every settled face. Those counts are compared against the distribution a genuinely fair pair produces, at a 0.01 significance level.

Rolls measured
10,000Real physics throws, not synthetic data
Verdict
FAIRAt α = 0.01
Totals χ² (2–12)
5.95 / 23.21Observed vs critical, 10 degrees of freedom
Per-die χ² (1–6)
4.81 / 15.09Observed vs critical, 5 degrees of freedom
Measured, 10,000 rollsExpected from a fair pair
04509001,3501,800234567891011121,713
Every total lands within about 2% of where a fair pair would put it. The biggest gap is on the 7, and even that is 46 rolls out of 10,000.
View the measured figures as a table
Measured versus expected counts for each two-dice total across 10,000 physics rolls.
TotalMeasuredExpectedDifference
2267278-11
3577556+21
4840833+7
51,0891,111-22
61,3571,389-32
71,7131,667+46
81,3831,389-6
91,0781,111-33
10854833+21
11551556-5
12291278+13

Both statistics come in far under their critical values: the totals at roughly a quarter of the threshold, the per-die at under a third. There were 0 settlement failures across the whole run: not one roll where a die landed cocked or failed to settle. The full methodology, the raw counts and the distribution charts are all in the report.

Don't take our word for it

Our sample is one dataset. Yours is another, and the apps ship the tools to check it: a fairness dashboard plotting your own lifetime roll distribution against the expected curve, for both dice totals and individual faces.

Export your entire roll history as CSV and run whatever analysis you like on it. If our dice were weighted, your own numbers would say so.

Fair dice, honest odds, and a house edge

Fair dice are not the same as a winnable game. Every bet on a craps table carries a house edge, and ours are the real ones: the field still costs 5.56%, any seven still costs 16.67%, and only the odds behind the line pay true.

That is the point of a simulator built this way. You can see exactly what the maths does to a bankroll over a long session, without it being distorted by dice that were never honest in the first place.