How we test

• base series one session thousands of ordinary rounds at a fixed bet, logged one by one; the exact count is on every test card
• bonus buys a series purchases at the listed price where the game sells its feature, same session
• media session once per game real browser, video with sound, recorded on the first run
• bet per spin 2.00 demo units the most common stake; many tests run at 1.00. Currency follows the demo server: USD for Pragmatic Play and Blueprint Gaming, FUN for BGaming and the Relax Gaming server, EUR for Nolimit City, Quickspin, Thunderkick and Hacksaw Gaming, DEMO for Play’n GO. Stake and currency are shown on each test card
• server official demo the provider's own game server
• retests over time the same game is run again as a fresh session; every run stays on the page with its own log

Every review on this site is built from one or more such runs. The rig talks to the game server the way the game client does: it opens a demo session, sends spin requests, reads each server response and records it. Nothing is simulated locally, no RNG is emulated. When the base series and the purchases are done, a real browser plays the same build for screenshots and video, and the numbers are computed from the log, not from what was seen on screen.

When a game is tested again, the new run is a fresh session on the same build, without new media. The headline figures of the review pool every run into one sample, as if they were one long session; the Test record on the review keeps each run apart, with its own return, interval, purchases and downloadable log, so a reader can see how far sessions of the same game drift.

So far: 163 tests · 2,820,000 logged rounds · 10,200 bonus buys · 18,807 natural bonuses · first run 2026-09-07, latest 2026-10-16.

1. 01 OpenThe rig requests the official demo build and receives a session key, the paytable, the bet levels and the RTP profile the server reports for this build.
2. 02 Base seriesOne uninterrupted session of ordinary spins at one bet. Each spin is a server request; cascades, free spins and the final collect are followed until the server credits the balance.
3. 03 Buy seriesA series of purchases of the feature at its listed price, same bet, same session. Games without a buy button skip this step.
4. 04 Media sessionA real Chrome window plays the build: rules screens, base spins, a bonus round captured from the tab at 1280×720 with sound. Recorded once per game, on the first run.
5. 05 MetricsRTP with an approximate 95% confidence interval, hit rate, distribution, droughts, drawdown, bankroll survival, buy value. All from the log.
6. 06 Checks and sign-offBalance reconciliation, text against log, page render on desktop and phone, then an editor reads the draft against the log. Nothing goes live before both stages are done.
7. 07 RetestsThe same game is run again later as a fresh session. The review pools all runs for its headline figures; the Test record lists every run with its own numbers, a dated lab note and its log with a checksum. The verdict of each run is written once and kept.

The archive also holds the raw server requests and responses and a separate purchase journal. Every run publishes, in the Test record of its review, a readable TXT spin log, an ordinary-round CSV and the gzipped JSONL of its rounds, purchases and raw server traffic, each with a SHA-256 checksum. The figures on the page are computed from these files, nothing else.

Measured RTP

Σ win ÷ Σ betTotal settled payouts divided by total stakes over the ordinary paid series, including naturally triggered features. Purchases are separate. This sample return is not the theoretical RTP or a fairness test.

Approximate 95% CI

mean(x) ± 1.96 · sd(x) ÷ √nApproximate uncertainty about the mean, using a normal approximation to per-round returns. It is not a prediction interval for the next run; rare large payouts make the approximation imperfect.

Pooled result

all runs in sequence, as one seriesWhen a game has several runs, their ordinary rounds are joined in order and every figure above is computed on that series, so the interval narrows as rounds accumulate. Each run keeps its own figures in the Test record.

Hit frequency

rounds with win > 0 ÷ roundsA round that pays anything counts, even 0.1x.

Bonus rate, bHz

natural bonuses ÷ rounds × 1,000A unit of this lab: one bHz means one naturally triggered feature per 1,000 ordinary rounds, so 5.2 bHz is a feature about every 192 rounds. Bought features are excluded. It is a rate, not a count, which is why a game tested over 10,000 rounds and one tested over 50,000 can be compared directly; the catalog sorts on it. Games whose base game pays nothing, where every round is already a feature round, carry a reels mark next to the figure.

Longest dry run

max run of consecutive win = 0Also reported: average dry run and the rounds where the longest one happened. Dry runs do not cross from one run into the next.

Distribution

0 · 0–1x · 1–2x · 2–5x · 5–10x · 10–20x · 20–50x · 50–100x · 100–500x · 500x+Share of rounds in each bucket of the bet multiple.

Max win, no bonus round

max(x) over rounds without a natural bonusThe full maximum including bonus rounds is shown next to it, with the run and round where it happened.

Drawdown

max(peak − net) ÷ betDeepest fall from the running peak of the balance, in bets, and the spin where it bottomed; with several runs the balance continues from one run into the next.

Bankroll survival

first spin where net ≤ −50 / −100 / −200 betsWhether a stake of that many bets would have been gone before the series ended, and when.

Buy value

mean(return ÷ price), median, share below priceComplete purchased-round payouts divided by the full ticket price. Mean, median and below-price count come from settled journal amounts without intermediate rounding. Extra-spin awards are defined by the particular game.

Latest 20 of 163 tests shown. Full register with sorting on the home page

Across 163 tests the measured return averages 95.9% against an average declared 96.38%. In 153 of 163 tests the declared figure sits inside the approximate 95% interval; this does not establish fairness; 90 tests came in below the declared value and the rest above. The lowest measured test is 70.5%, the highest 120.8%.

• Balance reconciliationFor every round, the logged win must equal the change in the server balance plus the bet. One mismatch means the protocol was misread and the run is discarded.
• Bonus accountingFree-spin totals are taken from the server's own end-of-feature fields and must match the per-spin sum.
• Every figure traced to the logEach number in the text (round numbers, multipliers, buy results) is looked up in the run before the page is built. After a retest, the series-level texts are rewritten from the pooled figures; the texts of earlier runs are not touched.
• No copied sentencesThe text of a new review is compared with every published one; shared phrasings are rewritten.
• Page renderStructured data, images, video and tables are checked on a desktop and a 390-pixel phone before publishing.
• Human sign-offAn editor checks the text against the log: every figure in the draft, the screenshots and the video. A review goes live only after this read.

Open the Test record of a review, pick a run and download its CSV as rounds.csv. The file covers the ordinary paid rounds of that run, including their natural bonuses; the TXT log also lists the separate purchases where tested. Recompute the measured RTP of the run with Python:

python -c "import csv;from decimal import Decimal as D;r=list(csv.DictReader(open('rounds.csv',encoding='utf-8-sig')));print(round(100*sum(D(x['win']) for x in r)/sum(D(x['bet']) for x in r),2))"

The output is the measured RTP of that run, in percent; it should match the figure on the run's plate. For a retested game, joining the CSVs of all runs reproduces the pooled figure on the test card. The square button next to each download copies its SHA-256 checksum, so the file can be checked for changes after download:

sha256sum rounds.csv

The result does not estimate what the next session will return. Hit rate and the largest complete round can also be checked from the CSV. If a figure differs, send the review link so the public export can be checked against the internal journal.