A shooter can set two dice to a chosen orientation. That part is observable. The disputed claim begins after the dice leave the hand: does the starting arrangement survive the flight, table impact, and rebound strongly enough to change the distribution of results?
There is no reliable reason for a player to price a craps wager as though dice setting alone creates an advantage. A controlled-shooting claim needs reproducible data from valid throws, not a memorable hand, a smooth-looking delivery, or fewer sevens during one session.
Setting is not the same as controlling
Dice setting means arranging the faces before the throw. Casinos commonly allow a shooter a reasonable moment to choose and align the dice, provided the game is not delayed.
Dice control or dice influence is the stronger claim that a repeatable throwing method changes outcome probabilities. That claim cannot be established by showing that the dice began on an axis. It requires evidence that the final totals occur at rates inconsistent with ordinary random variation.
This distinction prevents two opposite mistakes:
- “The casino lets me set the dice, so the method must work.”
- “The dice bounce, so influence is physically impossible under every condition.”
The first conclusion does not follow. The second is stronger than the evidence requires. The practical position is narrower: a player should not assume a usable casino advantage unless it has been demonstrated under realistic conditions.
Why a good-looking throw proves very little
Two fair six-sided dice have 36 equally likely ordered outcomes under the standard random model. Six combinations total 7, so:
P(7)=\frac{6}{36}=\frac{1}{6}\approx16.67\%The probability of avoiding a 7 for n consecutive rolls is:
P(\text{no 7 in }n\text{ rolls})=\left(\frac{5}{6}\right)^nExamples:
| Rolls without a 7 | Random-model probability |
|---|---|
| 10 | 16.15% |
| 20 | 2.61% |
| 30 | 0.42% |
A 20-roll hand feels exceptional, but it is not remotely impossible. Across many tables, shooters, and sessions, long hands will appear without anyone controlling the dice. Selecting the longest hand afterward and treating it as proof is a classic selection error.
The craps dice combinations page shows why totals are not equally likely, while craps probability basics explains how ordinary variation creates convincing short runs.
The correct test begins with a stated claim
“I'm better than random” is too vague. A test needs a measurable alternative.
Examples include:
- the shooter produces fewer sevens than 1/6 over point-cycle rolls;
- certain faces remain on axis more often than expected;
- pass-line wins occur more frequently than the random model predicts;
- hand lengths follow a distribution inconsistent with uncontrolled throws.
The measurement must be chosen before the data are reviewed. Otherwise, the tester can search dozens of outcomes and report whichever one looks unusual.
For a simple seven-rate test, let:
- n = number of valid observed rolls;
- S = number of sevens;
- \hat p=S/n = observed seven proportion.
Under the random model, the approximate standard error is:
SE=\sqrt{\frac{p(1-p)}{n}},\qquad p=\frac16With 100 rolls, random noise is large. With thousands of rolls, the estimate becomes more precise. A claimed edge must also be large enough to overcome the house advantage, not merely different from 1/6 by a tiny amount.
Conditions matter as much as the count
A credible casino-relevant test should document:
- identical, inspected dice or a recorded dice-change protocol;
- a normal table surface and back wall;
- valid throws under the property's rules;
- no discarded outcomes except predeclared no-roll conditions;
- an independent recorder or video record;
- a sample size chosen before the trial;
- the exact statistic and significance threshold;
- replication on another day and, ideally, by an independent team.
Throws that stop short, slide, strike chips, leave the table, use altered dice, or would routinely be called no-roll do not establish an advantage available in ordinary casino play. A laboratory effect that disappears under legal table conditions would not support a practical betting claim.
What published work has found
A 2020 original research article used a purpose-built throwing machine, high-speed video, and 7,557 recorded throws to test commonly described control methods. The researchers reported their experimental design and statistical testing in Pair-a-Dice Lost: Experiments in Dice Control. A machine is not a human shooter, but it is a demanding test of the claim because it can repeat release conditions more consistently than a person.
More recent work has focused on how a valid test should be constructed. Stewart Ethier's 2025 paper, Testing for Dice Control at Craps, compares statistics based on seven frequency, pass-line wins, hand length, and likelihood models. That is the right level of discussion: define the model, collect enough observations, and test whether the data support a meaningful degree of control.
Neither a published method nor a statistical test proves that every shooter is random. It shows what evidence would be needed before a claim deserves confidence.
The casino does not need to settle the physics debate
The property has a simpler operational concern: every throw must comply with house procedure and be easy to observe and call. The stickperson and boxperson watch whether both dice are thrown together, travel properly, strike the required area, remain on the table, and produce a readable result. Exact procedures vary.
A shooter who repeatedly creates short rolls or questionable contact may be warned, asked to change the delivery, or have a throw called no-roll. That response is game control, not proof that the casino fears a mathematically successful dice setter.
The dice-handling rules page covers valid-throw procedure. The separate dice control myth page examines the broader advantage claim.
How players accidentally manufacture evidence
Several habits make random play look controlled:
- counting only point-cycle rolls and forgetting come-out sevens;
- recording successful sessions but not failed practice sessions;
- changing the set or target after losses;
- excluding ugly throws after seeing the result;
- measuring multiple outcomes and reporting only the best one;
- comparing one shooter with a vague memory of “normal” results;
- stopping the test during a favorable run.
These are not small bookkeeping issues. Each one biases the sample toward the desired conclusion.
A practical decision rule
A player does not need to prove that physical influence is impossible. The betting decision has a lower burden: do not spend money as though an edge exists until repeatable, independently checkable evidence shows that it does.
Set the dice if the ritual is enjoyable and the table permits it. Keep the throw prompt and valid. But price every wager using the published game probabilities, not the appearance of control. A smooth release may improve comfort and consistency; it does not, by itself, rewrite the 36-outcome model.