How a maths model is verified
A maths model is a claim, and certification is the process of independently checking it before it reaches players.
An independent testing house (in Britain, one on the Gambling Commission's published list of approved test houses) receives the par sheet, the game software including its source code, and the supporting documentation, and does several distinct things.
Recomputes the RTP. For enumerable portions, independently. For simulated portions, by running its own simulation, over a number of plays set by the game's volatility, and checking that the actual RTP falls within an acceptable range of the expected figure.
Verifies the implementation matches the model. This is separate from checking the maths. A correct par sheet implemented incorrectly produces a game that does not behave as certified, which is why Britain's testing strategy lists verifying the software implementation of the game design, maths and theoretical RTP as a step separate from verifying the design itself.
Checks the paytable against the help screens. The rules shown to the player must match the rules the game applies: in Britain, games must be implemented in accordance with the rules as they are described to the customer. Mismatches here are a consumer protection issue rather than a mathematical one.
Tests the random number generator. Statistical randomness, unpredictability, non-repeatability, and seeding and scaling that do not introduce predictability or bias. A classic defect sits not in the RNG itself but in the mapping from RNG output to reel position, where modulo bias or incorrect range handling can skew the distribution even though the underlying generator is sound. Reducing a random number to a smaller range with the modulo operation leaves a skew, which only becomes negligible when at least 64 extra random bits are used, while discarding out-of-range values leaves none, and GLI-19 requires all scaling and mapping algorithms to be unbiased, as verified by source code review.
Checks the maximum win and truncation behaviour. A cap must work as stated, and the artwork must disclose maximum win values.
Confirms the configuration set. Where a game ships with several RTP settings, the testing needs to cover the settings that will actually be offered: GLI-19 requires the configuration that will be used in production to be communicated to the independent test laboratory.
Simulation confidence, and the question to ask
Where RTP is established by simulation, the declared figure has a confidence interval, and a rigorous maths sheet states it.
The principle follows directly from the volatility lesson: the standard error of a simulated mean falls with the square root of the number of runs. A game with a standard deviation of 15 per unit staked, simulated over 100 million spins, has a standard error of 15 divided by the square root of 100,000,000, which is 15/10,000 = 0.0015, or 0.15 percentage points of RTP.
That is acceptable precision. Now consider a hypothetical game with an extremely rare, extremely large top outcome. If that outcome contributes 2 percentage points of RTP and occurs once in 50 million spins, a 100 million spin simulation will have hit it roughly twice. The estimate of its contribution from two observations is close to worthless, and the overall RTP estimate inherits that uncertainty.
The correct treatment is to enumerate the rare high-value outcomes analytically and simulate only the remainder. Test houses work the same way: alongside simulation, Britain's testing strategy describes emulation testing, used to replicate rare outcomes such as jackpot triggers, special features and maximum prizes. The question to ask of any simulated model is therefore: which parts were simulated, over how many runs, and what is the stated confidence interval? A model that cannot answer has not been finished, whatever the certificate says.
RTP configuration and what legitimately changes it
Many titles ship with multiple certified RTP settings, for example a band such as 96%, 94% and 92%, and the operator selects one at integration.
Three things are true about this and are worth separating carefully.
It is legitimate. Each configuration is certified, and the practice exists because tax rates, competitive conditions and regulatory requirements differ by market. Malta, for example, requires online RNG games to pay out on average at least 85% of the amount wagered, or any higher figure set in the licence.
It must be disclosed. Regulators require the return to be disclosed: in Britain the RTP, the house edge or the probability of winning must be easily available before the customer gambles, and GLI-19 lets a regulator require the RTP in each game's artwork. The figure shown must be the one configured, precisely because the same game name can carry different returns at different sites. An operator running a lower configuration without disclosing it is the problem case.
It is per integration, not per player. The configuration is set for the operator's deployment. It must not vary by player, session or balance: in Britain adaptive behaviour, that is, a compensated game, is not permitted, and certification checks for it.
A related question that arises commercially: can an operator change the configuration later? Usually yes, subject to the supplier's process and to disclosure. In Britain such a change should be made with the game offline, and the altered game should display a notice such as 'rules changed' or 'different payouts', and GLI-19 has the operator treat each change to a game's theoretical RTP as a new game in its reports and records. What it cannot do is change it dynamically in response to play, and any system offering that would not be certifiable.
Reading a spec sheet like an operator
The commercial skill this course builds is looking at a game specification and knowing what you are being offered. The checklist:
RTP, and the configuration offered. Which bands are available, what does the supplier default to, and what does the competition run?
The base/feature split. Feature-led or base-led, which predicts most of the player response.
Volatility, as a standard deviation if you can get it. Star ratings are not comparable across suppliers.
Hit frequency, both headline and above stake. A wide gap between them means most hits are net losses.
Feature trigger frequency and average feature value. Together these describe the rhythm of the game.
Maximum win and its probability. Tells you whether the headline number is real or decorative.
Bonus buy price and buy RTP, where offered, and whether the buy RTP differs from base.
Jackpot contribution, where applicable, and the base RTP net of it.
Simulation confidence, for anything with cascades, uncapped multipliers or a very long tail.
Nine numbers. A supplier confident in its maths should be able to provide all of them to an operator that asks, and they are worth asking for before taking the game.
Building a portfolio with the maths
A game catalogue is a portfolio, and the mathematics gives you the axes to manage it on.
Volatility mix. Players differ in their appetite for volatility, and a catalogue weighted entirely to high-variance titles risks losing the audience that wants a long session on a modest bankroll, and vice versa. A study of 4,281 regular UK online slots players found the link between game volatility and behaviour complex, often non-linear, and different between a casino-focused and a bingo-focused operator's players. Segment the catalogue and check the distribution rather than assuming it.
RTP positioning. Running a lower configuration raises theoretical margin per unit staked and reduces session length, which reduces total staked. The net effect is not obvious and is empirically testable. An operator that lowered RTP across a catalogue and saw revenue fall has learned something real about its audience.
Speed and revenue per position. Since expected revenue is amount staked multiplied by edge, and amount staked is driven by speed and session length, a slow high-edge game can be worth less per position per hour than a fast low-edge one. Land-based floors have optimised this for decades; online catalogues frequently have not, because position is not scarce in the same way, but attention is.
Feature-led versus base-led balance. A catalogue of feature-led titles produces long dry spells for players who do not reach features, which shortens sessions for exactly the low-spending segment that provides volume.
Jackpot participation. Progressive titles draw participation as pools grow, at the cost of a worse base-game experience and noisier reported performance.
Convergence discipline. The volatility arithmetic says short-run performance data on a high-variance title is mostly noise. Set the observation period from the maths, not from the reporting calendar, and aggregate before judging. Britain's Gambling Commission makes the same point for live RTP monitoring: measurement frequency should be based on the volume of play, and volatility is a key parameter in setting each game's tolerance.
The habits worth keeping
Four, and they generalise beyond slots.
Never compare games on RTP alone. Expected loss per hour is the comparable figure, and it needs edge, stake and speed.
Always ask what the variance is before reading a result. A number without its dispersion is not evidence.
Ask which parts of a model were simulated. The answer separates a rigorous supplier from a confident one.
Remember that the maths describes the long run and the player lives in the short one. That gap is not a flaw in the mathematics. It is the reason good disclosure explains volatility as well as return, and it is the single most useful thing a maths-literate person can explain to a colleague who is about to say something misleading in public.