99.9720% is a full count, not a sample
The 99.9720% figure is the sum of every payout in a 2^32 space, divided by 2^32. That is 4,294,967,296 spin indexes. Nothing was skipped to save time. The weights and the paytable are the operator's published ones. An operator RTP label is not an input. A second implementation, CPU against a CUDA run, matched inside a stated float noise floor. Read 18 October 2026.
A spin index becomes a grid in two steps. HMAC-SHA256, with the server seed as raw hex bytes and a message built from the client seed and the nonce, produces an outcome index. PCG32, seeded from that index, then draws the symbols. Cascades use the next tumble index on that same stream. A matching grid can be rebuilt from the server seed, the client seed, and the nonce.
The bonus is not one spin, so it was not left to a sample. The session is a set of lanes: retriggers used, spins left. Each step folds the enumerated bonus payouts into the running total. Mass that would pass 5,000x is dropped into the cap bucket. The mean of what remains is the session's expected value. Turning the cap off moved the bonus return by about 0.0025 percentage points, so the 99.9720% is not an artifact of the lid.
The 30 million spin run is a check, not the source. It printed 100.06%. If that file is missing, the count still stands. You do not need to replay four billion spins to see the headline: the published test recomputes it from the stored enumeration.
Questions
Is 99.9720% a simulation?
No. The simulation is the 100.06% from 30 million spins. 99.9720% is the full count.
What goes into the count?
Published symbol weights and the paytable. Not a label that says 100%.
Can a spin be rebuilt?
Yes, from the server seed, the client seed, and the nonce, for grids that are inside the provably fair chain.