Naka-pin na Tweet
DUEL 100% RTP CASTLE ROULETTE AUDITED
👉 data sets available for download
Back in December I started recording every single Castle Roulette spin into my own database. Originally just to power the live stats page on rawcrew. Polled every 8 seconds, kept it running.
Five months later I looked up and I had 318,981 rounds sitting there.
At some point I realized this was not just a stats feed. It was an audit waiting to happen.
I told Duel what I was doing. @Dollartree_1 sent me 400,000 rounds straight from their backend. Raw, no filter, no cherry pick.
I crossed both datasets. Same game, same period, two completely separate sources.
What I found:
Out of 318,458 comparable rounds, 318,414 matched perfectly. That is a 99.99% match rate between two independent sources.
Every bet type sits at or very near the expected 100% RTP. 2x at 100.17% on my side, 100.12% on theirs. 4x at 100.04% on both. 8x at 100.04% on both. 16x at 99.29% / 99.40%. 24x at 99.49% / 99.57%. 48x at 98.46% / 98.99%.
Now if you saw those last three lines and thought "hold on, those are below 100%, something is off," that is a fair reaction. I want to address it directly because it is the first thing a careful reader notices.
Here is why those numbers look like that. Castle Roulette has 48 positions on the wheel. The 48x multiplier sits on exactly 1 of those 48. So in 318,000 rounds it is only expected to hit around 6,600 times. The 24x sits on 2 slots, the 16x on 3. Rare events on a small sample swing more, that is just how variance works. The lower the hit rate, the wider the natural spread around 100%, even when nothing is wrong with the game.
If the sample was 10 million rounds those numbers would tighten right back to 100%. With a few hundred thousand they wobble a bit. That wobble is expected, not suspicious.
The statistical tests confirm it. Chi-squared goodness of fit at 2.63, well below the 11.07 threshold. P-value of 0.76. A perfectly fair game would produce this much variation or more 76% of the time. It passes easily.
Runs test for randomness: 200,722 runs, inside the expected band. Passes.
Repeat analysis across all six multipliers came back inside expected bounds. No clustering, no manipulation.
The wheel mechanic itself is textbook provably fair. Server seed committed before the round, drand beacon as external randomness neither side controls, HMAC-SHA256 to combine them. Anyone can replay any round and verify it. 0% house edge confirmed.
Now the part that matters most. I am not a research firm. I am one guy pulling data and doing my best to read it honestly using lots of AI. I could have made mistakes.
That is exactly why I am sharing both raw datasets, mine and Duel's. Link in the comments. Download them, open them up, run your own numbers. If I got something wrong I want to know.
Thanks to Duel for sharing their official data. That kind of transparency is what made this even possible.
English
