Dice, Coins & Card Randomisers

Coin flip: the cleanest possible binary draw

Two outcomes is the one case where the byte-to-outcome math needs no cleverness at all -- half of every possible byte value is even, half is odd, so checking that single bit hands back a dead-even 50/50 split without any discard-and-redraw step. The interesting comparison isn't to the math, though, it's to an actual coin: mathematician Persi Diaconis's research on physical coin tosses found a real coin tends to land back on whichever face it started on around 51% of the time, a small lean caused by how the mass sits on either side. A digital flip was never sitting on a table to begin with, so that particular quirk has nothing to attach to.

Dice roller: exact per-face fairness for tabletop games

A 20-sided die is the clearest example of why this site bothers with rejection sampling at all: 20 doesn't split cleanly into the 256 values a random byte can take, leaving a lopsided remainder that would quietly favor the lower half of the die's faces if you mapped bytes onto numbers the easy way. The fix is to throw away any byte landing in that remainder and pull a fresh one, over and over if needed, until every face from 1 to 20 is equally reachable. Tabletop games lean on that exactness more than people usually realize -- a natural 20 is supposed to be a 1-in-20 event, not a 1-in-17 event dressed up to look the same, and that gap only shows up statistically after enough sessions to notice.

Manufactured dice have their own quiet failure mode here too: a die that's ever so slightly heavier on one face, or has a corner rounded a touch more than the rest, will drift off true fairness in a way no player could spot by eye. A byte-corrected digital roll never has a heavy side to worry about.

Magic 8-Ball: a faithful reproduction, not a clean 50/50

This is the odd one out in the category, on purpose. Where the coin flip and dice roller are both trying to be as flat and even as possible, the 8-ball is trying to be faithful to a specific physical toy that was never flat to begin with -- 20 printed replies split 10 affirmative, 5 hedging, 5 negative. The draw across those 20 lines borrows the same discard-and-redraw approach the dice roller needs for its 20 faces, but the goal here isn't neutrality; it's accuracy to the original object. Ten of twenty replies leaning positive means a 'yes'-shaped answer really is somewhat more common than a 'no'-shaped one, and that's correct behavior for this tool rather than a bug to fix.

Choosing between coin, dice, and 8-ball

For a clean, genuine two-outcome decision, the coin flip is the right tool -- exact 50/50, no ambiguity, no extra flavor. For anything that needs more than two outcomes with exact per-face odds, especially tabletop gaming where the specific face-count math matters (a d20's 1-in-20 crit chance, for instance), the dice roller is purpose-built and supports the standard d4 through d20 range. For a nostalgic, deliberately non-committal answer to a yes-or-no question -- where you're not actually looking for a decisive verdict, just a bit of fun -- the Magic 8-Ball's 20-answer set, including its 'ask again later' style hedges, is the right choice specifically because it isn't a clean binary.

The gambler's fallacy, and why streaks don't change the odds

A common misconception across the coin flip, dice roller and Magic 8-Ball is the gambler's fallacy: the belief that after a run of the same outcome (say, five coin-flip heads in a row), the opposite outcome becomes 'due.' It isn't. Each flip, roll, or 8-ball draw on this site is a fully independent random event -- the tool has no memory of previous results, and the random-bit or rejection-sampling mechanism underneath doesn't track history at all. A coin that's landed on heads five times in a row still has exactly a 50% chance of heads on the sixth flip, precisely because the sixth flip is a completely fresh, independent draw with no connection to the previous five.

This is worth naming explicitly because it's genuinely counterintuitive -- human pattern-recognition instincts push hard toward expecting 'balance' to reassert itself, but true independent randomness doesn't work that way, and a tool that artificially corrected for streaks to 'feel' fairer would actually be less fair, not more.

What the coin, die and 8-ball do not claim to do

It's worth being explicit about the limits here. None of the tools in this category are built for gambling with real money, and none of them include any mechanism for verifying a result to a third party after the fact beyond what you can see and record yourself in the moment. They're built for the everyday, low-stakes version of these classic decision objects -- settling a chore dispute, running a tabletop game, having a bit of nostalgic fun with an 8-ball -- not as a certified random-number source for a regulated use case, which would need a different kind of auditability than a simple client-side tool provides.

Reading results with a critical eye

One useful habit when working with any of these tools for something that actually matters -- splitting a real prize, deciding a real chore rotation over weeks -- is running a larger batch and checking the overall distribution rather than trusting a single draw's 'feel.' Flip the coin 50 times and count heads versus tails; roll the die 100 times and chart how often each face comes up. Because the underlying draw is genuinely uniform, a large enough batch should land close to the expected even split, and a batch that looks meaningfully skewed is worth treating as a signal to double-check rather than dismiss, the same instinct a careful person would apply to a physical die or coin they suspected might be loaded.

This same batch-testing instinct is exactly how researchers like Persi Diaconis originally documented the real physical coin-flip bias mentioned earlier -- not by suspecting a single flip, but by tossing thousands of coins mechanically and measuring the aggregate result. A digital tool built on crypto.getRandomValues should, and does, hold up to that same scrutiny far more cleanly than a physical object ever could.

Counting successes rather than adding faces

The dice pool roller asks a different question from the standard dice roller and gets a different distribution as a result. Summing dice produces a bell curve that tightens as you add more; counting how many beat a target produces a binomial spread that stays wide.

Six ten-sided dice needing seven or higher succeed forty percent each, so the expected count is 2.4 — and zero successes happens about 4.7 percent of the time, roughly one roll in twenty-one. A summed roll of six dice essentially never produces its minimum, which is why pool systems feel swingy and sum systems feel predictable from the same dice.

Exploding dice change the model rather than adding a bonus: a maximum face rolls again and the extra die joins the pool, so a pool of six can return seven successes. The expected pool size becomes n divided by one minus one over the number of sides, which for six ten-sided dice is 6.67 rather than 6 — a real increase, and considerably smaller than the drama of an explosion suggests.

Faces you write, and a deck you shuffle

The custom dice builder is the same cumulative-sum weighted draw the picker category uses, applied to faces you type. The important thing it adds is the probability table, and the reason is specific: an unequal spinner wheel shows its odds as visible arc area, and a loaded die shows nothing at all. Printing the percentages is the only available substitute for the visual cue a wheel gets for free.

It also removes a constraint physical dice cannot escape. Real dice are limited to shapes that can be made fair, which is why there is no fair three-sided die; software has no such limit, so a three-sided, seven-sided or nineteen-sided die all work identically.

The deck shuffler performs a single Fisher-Yates pass over fifty-two cards, which unlike a physical riffle does not improve with repetition — after one complete pass every ordering is already exactly equally likely. The number worth pausing on is 52 factorial, about 8 followed by 67 zeros: the ordering you just produced has, with overwhelming probability, never existed before and never will again.

Measuring what randomness actually looks like

The streak simulator exists to answer a specific recurring argument: somebody has seen a run of six and concluded the tool is broken. Running two hundred simulated flips in front of them and getting a run of seven, from a generator nobody has touched, is faster and more convincing than any explanation of independence.

The theoretical figure it prints alongside is the base-2 logarithm of the flip count, which is the standard approximation for the expected longest run. Its slow growth is the counter-intuitive part: two hundred flips gives about 7.6, and twenty thousand gives about 14.3, so a hundredfold increase in flips adds under seven to the streak.

Running several batches at once surfaces the variance that a single number hides. Ten batches of two hundred flips typically produce longest runs from about five to about eleven, which makes the case that a single session cannot tell you anything about a generator without anyone having to argue it.

Frequently asked questions