How the fairness monitor works
Every check compares our own recorded rounds with a figure the provider or operator publishes, gives a 95% range for what we measured, and turns the gap into a status by rules set in advance. Nothing is set by hand.
What a status can and can't tell you
"Consistent" means our rounds give no statistical reason to doubt the published figure. It is not proof that a game is fair: a small edge can hide inside any range, and the narrower the range, the more rounds it took. "Not consistent" means a gap larger than chance explains after correcting for every check on the monitor. It says the recorded results and the published figure don't match; it does not say why.
Method 1: crash and slide (12 games)
Each round ends at one multiplier. If the operator's formula promises a return factor r, the share of rounds that reach a target m is r divided by m (adjusted for the operator's rounding and any fixed share of instant 1.00x rounds). We check three things:
- The shape of the curve. A chi-square test on how the results above 1.00x spread across eleven bands, against the formula. If the shape is off, the next check is not judged.
- The implied return. A maximum-likelihood estimate of r from every round. Under the published formulas the curve's shape is fixed, so this estimate rests on the share of rounds that end at 1.00x. It is the most precise single figure we can give, and it only counts when the shape check passes.
- The return at each target (1.5x, 2x, 5x and 10x, then every step of 10x from 20x to 100x): target times the share of rounds that reached it, with a Wilson 95% range, each judged on its own against the formula. When a range is wider than plus or minus 1 point we show "too little data" instead of a number; a target above 10x is left out until its range is narrow enough, and appears by itself as rounds build up. Precision drops as the target rises: at 2x, about 150,000 rounds give a range of plus or minus 0.5 points; at 20x it takes about 720,000 rounds to reach plus or minus 1 point, and at 100x about 3.7 million.
Where the operator's formula is not public, there is no curve or estimate to test. Each target is then checked directly against the published RTP.
Verified rounds. For operators that publish a hash chain, a seed or a random value per round, every result is recomputed with the operator's own formula before it is stored. Verification shows the results follow the formula. It is not a fairness test; the checks above are.
Method 2: live game shows, roulette and dice (23 games)
- How often each result lands, against the published wheel or table layout: a chi-square test across all results, plus each wheel result on its own. Bonus rounds are tested this way only, because several of them pay each player a different amount.
- The RTP of fixed-payout bets, where our recorded rounds price the bet exactly: the mean return per 1 staked, with a 95% range from 2,000 resamples of the rounds (a large bonus multiplier makes the distribution lopsided, which a plain normal range handles badly in short timeframes). A published range is tested against its nearest end.
Bonus rounds that pay each player a different amount are priced from what was actually paid: a random pick on a Cash Hunt board is worth the board's mean, a Crazy Time flapper the mean of the three.
The thresholds
- Every check gives a two-sided p-value.
- All checks on the monitor are corrected together with the Holm method, which keeps the chance of even one false alarm across the whole monitor below the threshold.
- Not consistent: corrected p below 0.01.
- Borderline: uncorrected p below 0.01, but not "not consistent". With many checks, a few of these are expected by chance alone.
- Too little data: fewer than 1,000 rounds, an RTP range wider than plus or minus 2.5 points, or a chi-square with any expected count below 5.
- Short timeframes on a game's page are corrected within that page only.
We test the method itself on simulated rounds drawn from known distributions: the 95% ranges cover the true value about 95% of the time, and with the Holm correction a family of 20 fair checks raises a false alarm in well under 8% of runs, against more than half without it.
Dice games
Dice need no published table: fair dice determine every figure. Three dice give three of a kind in 1 round in 36 (2.78%), each face on 1 die in 6, and the even-money Big and Small bets win 105 rounds in 216 each, a return of 97.22%. In Bac Bo, a side rolls a pair 1 time in 6, a tie comes up 146 times in 1,296 (11.27%), and Player and Banker win 44.37% each. We measure each of these on every recorded round.