Coin Flip

A coin flip that looks convincing on screen and one that is actually fair are two different engineering problems, and it is entirely possible to get the first without the second. The straightforward-looking approach — spin a coin graphic for a plausible duration and read off whichever face is showing when it stops — makes the outcome depend on animation timing and easing curves rather than on a genuine random draw, and subtle implementation details can skew that toward one face more often than the other without anyone noticing from casual use. This tool decides the outcome first, using `crypto.getRandomValues` to draw a single unbiased bit, and only then plays a flip animation that lands on the result already chosen — the animation is pure presentation with zero influence over which face comes up.

Human intuition about randomness is systematically miscalibrated in one specific way: people expect a fair process to alternate more often than true randomness actually does, so a run of several identical results in a row reads as evidence something is wrong even when it is exactly what unbiased randomness produces. The probability of five heads in a row from a fair coin is 1 in 32 — unlikely for any single specific run, but routine across enough flips, since with a few hundred flips a streak of that length or longer is close to guaranteed to show up somewhere. This is the same cognitive pattern behind the "gambler's fallacy," the mistaken belief that a run of one outcome makes the opposite outcome "due" — each flip is fully independent of every flip before it, so a coin that just landed heads five times in a row is exactly as likely to land heads a sixth time as it was on the first flip.

Research on physical coin flips has found that a coin genuinely tossed into the air and caught is extremely close to a fair 50/50 process, because the number of rotations before landing is effectively unpredictable at the timescales and forces involved. A coin spun flat on a table or a hard surface behaves very differently: its center of mass is very slightly off-center toward the heavier, more detailed face on most coins, and that asymmetry biases a spun coin to settle showing the lighter face more often than chance would predict, by a margin some studies have measured at around a 1 percentage-point tilt or more depending on the coin. A cryptographically random digital flip sidesteps the entire question, since there is no physical object with an off-center mass to introduce a bias in the first place — the outcome is a mathematically unbiased coin toss every time, which is a stronger and more verifiable guarantee than even a carefully tossed physical coin can offer.

A single flip only ever answers one question — heads or tails, this time — and says nothing about whether the underlying process is actually fair, since even a badly biased coin will occasionally land on its less-likely face. Running a large batch of flips is what actually demonstrates fairness: as the number of flips grows from 10 to 100 to 10,000, the aggregate heads/tails split reliably converges toward an even 50/50 ratio if and only if the underlying source is genuinely unbiased, which is the law of large numbers in action rather than a coincidence. This is also the practical reason statisticians and quality-assurance engineers test a random source with large batches rather than a handful of trials — a small sample can look fair or unfair purely by chance, but a large enough batch cannot hide a genuine bias for long.