methodology

what this lab does

slots·science simulates slot sessions and publishes what happened. we build statistical models of slot games, run them millions of times, and report the outcomes, bust rates, session lengths, bonus waits, final-bankroll distributions, with confidence intervals. that's the whole operation. we don't review games, we don't rate casinos on vibes, and we never tell you to bet.

what we simulate, and what we don't

we simulate models calibrated to published math, not the provider's game engine. we don't have access to any provider's actual code, reel strips, or RNG, and neither does anyone else outside the provider and its testing lab. what providers do publish, return-to-player (RTP), volatility class, hit frequency, maximum win, bonus mechanics, feature-buy prices, is enough to build a statistical model that behaves like the game at the session level: how fast bankrolls drain, how often bonuses arrive, how outcomes spread.

this is a limitation and we state it in every study. it is also the reason the numbers can be trusted: a model calibrated to the published 96.50% RTP will show you what 96.50% does to a bankroll over a session, which is the question the published number alone never answers.

per-spin payout shapes (we typically use lognormal distributions) are model assumptions. distribution percentiles in our studies, "the median bonus paid 52.9x", are model-based estimates, and the studies label them as such.

how a model earns the right to be published

every model passes two validation gates before a single result is reported. the simulation script enforces both:

1. analytic check. the model's closed-form expected value must equal the target RTP to within 0.01 percentage points. calibration is exact by construction, so any gap means broken inputs, and the model is rejected.

2. simulation check. a 10-million-spin verification run must land within 4 standard errors of the target RTP. this verifies the implementation, not the math: heavy-tailed slot distributions can't be sampled to better than about ±0.4 percentage points even at 10 million spins. (remember that the next time anyone, including a casino, claims to have "measured" a slot's RTP from a few thousand sessions.)

a model that fails a gate is rejected and rebuilt from the inputs. it is never hand-adjusted until it passes, and it is never tuned to produce a more interesting result. if results look wrong, we investigate the model and document every change and the reason for it in the study's model file.

sample sizes and confidence intervals, in plain language

a typical study runs 9 stake/bankroll combinations × 10,000 sessions each, with a 2,000-spin cap per session, 10 million validation spins per model, hundreds of millions of simulated spins per study, and studies that test RTP variants or disputed parameters re-run the entire grid per variant.

• bust rates carry 95% confidence intervals (normal approximation: p ± 1.96·√(p(1−p)/n), matching what the simulation engine computes and records in each results file). at 10,000 sessions that's roughly ±1 percentage point. when we say "72.4% ±0.9", we mean: if the model is right, the true rate is inside that band 95 times out of 100.
• bonus statistics are only published if the bonus fired at least 1,000 times in sample. if it didn't, we widen the run or drop the claim.
• superlatives ("longest dead streak observed: 49 spins") are labeled as sample observations, not properties of the game. a bigger sample would find a bigger streak.
• rounding is honest. no decimal places beyond what the confidence interval supports.
• inputs are not findings. some outputs mechanically reproduce what we fed in, the average bonus gap echoes the trigger rate, the mean buy payout echoes the published buy RTP. we present those as model properties, never as discoveries. the findings are the things the inputs don't trivially imply: bust rates, survival shapes, distribution skew, streak tails.

and when a metric can't be supported by the sample, the study says "our sample cannot resolve this." we'd rather print that sentence than a shaky number.

where the inputs come from: the two-source data rule

inputs follow a fixed source hierarchy: provider-official pages first, then licensed-operator game pages, then independent aggregators and trackers as cross-checks. every input fact in a model, RTP versions, hit frequency, trigger rates, buy prices, max win, needs two independent sources before we treat it as established. where only one source exists (a single large-sample tracker, an unverified aggregator figure), the study says so plainly and the result is framed accordingly. where sources disagree, we don't pick a side quietly: we simulate both readings and publish both grids. the gates of olympus study brackets a disputed 1-in-209 vs 1-in-448 trigger rate exactly this way.

RTP-variant verification

many modern slots ship in multiple RTP configurations, the same game can pay 96.5% at one casino and 94.5% at another, with nothing on screen telling the player which version they got. we simulate the provider-default version unless stated otherwise, always state which version was simulated, and re-run grids at the published variants when the spread is material. every variant we publish is verified against at least two independent sources, with the provider default identified; single-source variant claims are footnoted as unverified, never counted, never headlined.

which casino runs which version is established from aggregator scans and tracker data, dated, and labeled as observations ("sighted at 96.50, june 2026"), operators can change configurations at any time, and a sighting is not a guarantee. where our affiliate links point to a casino, it is because that casino was last observed running the game's best published rate. that is a factual statement about configuration, never a claim that you'll win more, at every published RTP, the game remains negative-expectation.

what we never claim

• that any result predicts your session, your luck, or what the game will do next. simulations describe distributions; they do not foresee outcomes.
• that any slot, version, casino, stake, or strategy gives players an edge. none does. every slot we study is a negative-expectation game and every study says so.
• that a slot is "due", "hot", "cold", or beatable. spins are independent; the math doesn't care what just happened.
• that any result is a strategy that wins. lower bust rates at lower stakes mean losing more slowly, and our studies phrase it exactly that way.
• that we measured a provider's actual engine. we model published math (see above).
• any number we can't trace. every figure in a published study traces to a results file or a cited source. if we can't trace it, it doesn't print.

why we publish bust rates

a slot's marketing tells you the maximum win. the RTP figure tells you a long-run average no single session will experience. neither tells you the thing a player actually faces: how likely is this stake, on this bankroll, to hit zero, and how fast? that's what bust rates, survival curves, and final-bankroll distributions measure.

we publish them because they show the cost of play honestly, in the unit that matters (your bankroll), before you spend it. that is the lab's version of responsible gambling: not a slogan in the footer, but the actual price tag, measured, with error bars. if seeing the number makes someone play less, or not at all, the page did its job. see our responsible gambling page for help resources.

corrections policy

we will get things wrong eventually; the policy is what happens next.

• errors are corrected loudly, not quietly. a corrected study carries a dated correction note at the top of the page, what was wrong, what it said before, what it says now, and why it was wrong. we do not silently edit numbers.
• material errors (a wrong headline figure, a misattributed statistic, a broken model input) also get flagged through the same channels that distributed the original claim.
• the original results files stay archived; corrections add, they don't erase.
• found an error? email us. checking our math is the point, the data files behind each study are published so you can.

simulation today, real data next

everything above describes simulation: models of published math, run at scale. the next stage of the lab is real-data calibration, anonymized, per-game aggregate data from partner operators, used to publish "simulation vs. observed" comparisons and to calibrate models against actual play. the program's rules are already fixed: aggregates only, never individual player data; data sources disclosed in every study that uses them; and editorial independence locked in writing, no partner reviews, edits, or vetoes a result. until a study explicitly says it uses operator data, every number on this site is simulation-based, and labeled as such.

review before publish

no study is published as drafted. every draft passes an internal compliance review against the rules on this page, sourcing, framing, the forbidden-claims list, the confidence-interval discipline, before it goes live. it's a slower way to run a content site. it's the only way to run a lab.

affiliate funding, disclosed

slots·science is funded by affiliate commissions: if you sign up at a casino through our links, we may earn a commission at no cost to you. the funding never touches the numbers, simulation results are produced and locked before any affiliate placement happens, and our compliance review fails any draft whose figures can't be traced to the raw results. the business model is honest data with the monetization built around it, not the other way round.

18+. slots·science is for players in markets where online play is legal; we do not target the US or UK.