Poker, together.Back to Gather ↗A fresh deck
Each new hand creates a standard 52-card deck and shuffles it with Fisher–Yates, using Node’s cryptographic random-number generator for all 51 shuffle choices.
THE GATHER ENGINE
Fair cards.
Open results.
A good poker night starts with trust in the deal. Here is how Gather shuffles, what we tested, and the evidence behind the results.
Audit results consistent with a fair shuffle
Published September 19, 2026 · Gather engineering self-audit
1,000,000six-player deals tested
6,000,000player hands checked
0ranking or dealing failures
Evidence you can inspect.
We found no statistically significant deviations after correcting for 3,227 comparisons. An additional 200,000 deals supplied comparison estimates. No shuffle or poker-engine code was changed as a result of the audit.
This is a statistical audit of a specific engine version, conducted by Gather. It is not independent certification or a cryptographic “provably fair” system. No finite test proves that a random system can never fail.
Download the aggregate evidence ↓Read the methodologyFROM DECK TO SHOWDOWN
One engine. Every seat.
The same deal for everyone
Practice games and games with human players share the same shuffle and dealing path. Computer players choose actions using their own cards and the public board; they cannot see or replace the deck.
Best five cards win
The engine evaluates the best five-card hand from each player’s seven cards. We independently checked every category and every winning or tied player in the million-deal sample.
THE MEASUREMENTS
Observed meets expected.
Exact probabilities come from combinatorics. Where exact six-player calculations are impractical, we label the separate simulation estimate and its uncertainty.
| At least one player makes… | Observed | Exact expectation |
|---|---|---|
| A flush, including straight flush | 11.870% | 11.851% |
Board patterns
“Four to a straight” means four distinct ranks in any legal five-rank straight, including inside draws and ace-low straights.
| Board event | Observed | Exact expectation |
|---|---|---|
| Exactly one pair on board | 42.180% | 42.257% |
| Double-paired board | 4.783% | 4.754% |
| Exactly trips on board | 2.116% | 2.113% |
| Any repeated board rank | 49.255% | 49.292% |
| At least trips on board | 2.292% | 2.281% |
| Monotone flop | 5.177% | 5.176% |
| Exactly four board cards of one suit | 4.271% | 4.292% |
| Five board cards of one suit | 0.204% | 0.198% |
| Completed board straight | 0.380% | 0.394% |
| Four to a straight, including completed | 13.597% | 13.652% |
| Four to a straight, not completed | 13.216% | 13.258% |
Individual player hand categories
Six players share the same board. Statistical uncertainty accounts for that dependence instead of treating all six million player hands as independent.
| Best hand | Observed | Exact expectation |
|---|---|---|
| High Card | 17.446% | 17.412% |
| One Pair | 43.811% | 43.823% |
| Two Pair | 23.495% | 23.496% |
| Three Of A Kind | 4.834% | 4.830% |
| Straight | 4.590% | 4.619% |
| Flush | 3.024% | 3.025% |
| Full House | 2.598% | 2.596% |
| Four Of A Kind | 0.170% | 0.168% |
| Straight Flush | 0.031% | 0.031% |
Winning hands at six-player showdown
Each deal counts once, including ties. These comparison estimates use a separate 200,000-deal run of the production shuffle; they are not exact theoretical constants or an independent random-number generator.
| Winning hand | Observed | Comparison estimate · 95% interval |
|---|---|---|
| High Card | 0.051% | 0.053% · 0.044%–0.064% |
| One Pair | 16.231% | 16.140% · 15.980%–16.302% |
| Two Pair | 30.435% | 30.453% · 30.252%–30.656% |
| Three Of A Kind | 13.894% | 14.026% · 13.875%–14.179% |
| Straight | 16.569% | 16.544% · 16.382%–16.707% |
| Flush | 10.854% | 10.802% · 10.667%–10.939% |
| Full House | 10.906% | 10.966% · 10.830%–11.104% |
| Four Of A Kind | 0.890% | 0.856% · 0.816%–0.897% |
| Straight Flush | 0.172% | 0.160% · 0.143%–0.178% |
Strong hands arrive in clusters, too.
Trips or better won 53.284% of the six-player showdowns. The longest observed runs were 20 trips-or-better winners and 13 straight-or-better winners across four 250,000-hand streams. A surprising short sequence alone does not establish a shuffle defect.
These deals forced all six players through river. Real games include folds and eliminations, so their winning-hand mix can differ.
HOW WE CHECKED
The method, without the mystery.
- Test the deployed engine. Isolated processes ran the exact production image and Node 24.21.0 runtime. Every hand called the production start, shuffle, action, burn, board-dealing, evaluation and settlement functions. There was no substitute shuffler or supplied test deck.
- Keep the sample independent. Four fresh streams of 250,000 deals reset all six stacks each hand. Small equal stacks and legal calls forced every player to showdown. No cards, stacks or elimination state carried into the next hand. The extra comparison stream contained 200,000 deals.
- Measure more than winners. We counted every card, rank and suit; every deck position; board patterns; rank concentrations; player and winner categories; runs; and successive-shuffle correlations. The exposed sample comprises twelve hole cards plus the five-card board, excluding burns.
- Check with independent mathematics. An independently written rank-count and suit-run classifier checked the production evaluator. Card and board expectations use hypergeometric probabilities and exact enumeration. Seven-card category expectations use published combinatorial counts.
- Account for ordinary variance. Results include confidence intervals and statistical comparisons. Holm correction accounts for the thousands of tests. Overlapping streak windows and shared boards are treated as dependent. The largest individual position deviation was 3.42 standard deviations, unsurprising among thousands of comparisons.
- Check fresh randomness. A separate 1,000-hand observation probe verified 51 correctly bounded cryptographic calls per shuffle, fresh deck arrays and no adjacent deck repetition. No repeated deck appeared within any million-sample stream. Practice Mode source review found no deterministic fixture or deck override in the new-hand path.
Node may buffer unused cryptographic random bytes. Every draw consumes fresh bytes; buffering is not reuse of a shuffle, seed or previous output. Saved hand decks support replay of an existing hand, not generation of a new hand.
Exact formulas and definitions
- A specific card appears among the 17 exposed cards with probability 17/52; at any particular deck position its probability is 1/52.
- Rank and suit counts follow C(K,k) × C(52−K,n−k) / C(52,n), with K=4 for a rank or K=13 for a suit.
- Board rank patterns enumerate all five-card rank multisets, weighted by the product of C(4,m). Their total is C(52,5)=2,598,960.
- A monotone flop has probability 4 × C(13,3) / C(52,3).
- All four of a specified rank among 17 cards: C(48,13) / C(52,17). Two specified complete ranks: C(44,9) / C(52,17). “Any two ranks” is different; the exact calculation avoids double-counting deals with more than two complete ranks.
- The exact concentration distribution is the coefficient of x¹⁷yᵗzᑫ in (1+4x+6x²+4x³y+x⁴z)¹³, divided by C(52,17).
- For an independent per-hand event with probability p, a specified k-hand consecutive block has probability pᵏ. Searching anywhere in a longer sequence changes that probability.
- Contains-flush includes straight flushes and hands whose stronger best category supersedes a flush. Exact best categories are listed separately.
Scope and version
This dated audit is a snapshot, not a continuously running guarantee. It does not test every possible operational failure or certify operating-system entropy. Six-player comparison simulations use the same production shuffle; source review, exact mathematical checks and independent hand classification supply separate checks.
No private production player cards, accounts or hand histories are included in the public downloads. The evidence consists of synthetic aggregate measurements.
Audited shuffle module SHA-256:
95f51b4d8bd99991c47e6554ee5c09a1d56b8c6eca8324679ae8e0292ba746d4Audited dealing / Hold’em module SHA-256:
e9b1b538b6149ede389b3aae42afaf8fce897055e89ae8390ae633a97b8b3328CHECK THE NUMBERS
The evidence is yours to read.
CSV files contain observed counts, expectations and statistical comparisons. JSON samples contain aggregate histograms, transitions, run counts and timing metadata. The archive includes a SHA-256 integrity manifest.
Every card CSV ↓Every deal position CSV ↓Rank frequencies CSV ↓Suit frequencies CSV ↓Board / exposed rank and suit distributions CSV ↓Board patterns CSV ↓Player categories CSV ↓Winning categories CSV ↓Six-player made-hand probabilities CSV ↓Exact six-player flush check CSV ↓Rank concentrations CSV ↓Full streak distributions CSV ↓Consecutive qualifying hands CSV ↓Successive-hand correlations CSV ↓All 3,227 corrected statistical checks CSV ↓
Download all aggregate evidence ↓
Primary references: Brian Alspach’s seven-card hand counts · Node 24.21.0 cryptographic randomness implementation