# Statistical audit of an electronic roulette draw series — interpreting daytime vs night biases

I'm doing a statistical audit of the public draw history of an electronic roulette lottery game (Ruleta Express, Dominican Republic). Looking for advice on randomness-quality tests and how to interpret structured biases — not gambling advice, and I'm not asking anyone to bet.

## Game shape (public)

- 19 pockets, numbers 0–18. Draw about every 5 minutes. Pays 18×.
- Historic operator was Intralot/LOTOS until ~2020; current RNG implementation is not public.

## Observed patterns in the history

- **Daytime window (~07:38–19:59):** very strong anti-repetition — a number almost never repeats the previous draw (~0%).
- **Late night:** repeats around ~5% (closer to uniform independent draws).
- **"0"** appears roughly ~10% less often than a fair 1/19 would predict, across the day.

## What I already checked (no credible match)

- Classic daily-seeded 32-bit PRNGs against the series: MSVC/ANSI rand, glibc rand(), Java LCG, .NET Framework LCG, MT19937.

## Questions for people who audit RNG / randomness quality

1. Does a daytime anti-repetition rule layered on top of an otherwise uniform generator fit this better than a single biased generator? How would you separate those hypotheses statistically?
2. Sensible next randomness tests on a categorical 0–18 series: PractRand / TestU01 (after a suitable encoding), chi-square / runs / autocorrelation, spectral / period checks, or something else cheap?
3. For lottery-style electronic draws, are scheduled rule layers (different daytime vs night behavior) common in certified systems, or more often a symptom of a non-certified path?
4. Any references on auditing short discrete alphabets (19 symbols) rather than bitstreams?

Happy to share anonymized summary stats. Again: analysis / audit only — please don't bet on this.

---

*Discussion thread: https://www.reddit.com/r/cryptography/comments/1xsxa60/statistical_audit_of_an_electronic_roulette_draw/*
