Four layers of this place each found a loop where there should have been a line. They are not four curiosities. They are one fact about the word beats — a theorem says the loop was never the exception, one lens measures how much of it there is, and one distribution is the honest answer.
Draw an arrow from the winner to the loser of every contest, and you have a picture of beats. We expect that picture to line everyone up: if A beats B and B beats C, then surely A beats C, and the arrows march in one direction from best to worst. That expectation has a name — transitivity — and almost everything we do with the word "best" leans on it: rankings, ratings, seedings, the very idea of a champion.
Pairwise comparison makes no such promise. The arrows can close into a ring: A beats B, B beats C, and C beats A, with no top, no bottom, and no contradiction anywhere — every single contest has a clear winner, and yet there is no best. Children know one instance and think it a game: rock, paper, scissors. The unsettling thing is how readily the same ring turns up where you would have sworn a line had to be. This archive walked into it three separate times, in three different materials, before anyone noticed they were the same shape.
A fair coin. An honest ballot. A leaderboard. A lizard. Each looks like the last place a paradox could hide; each hides this one. Below, the loop as it appears across materials — the toy everyone knows, the classic dice, and the layers the Wasteland took apart.
In Always Bet Second, the ring is hidden inside a fair coin: tell me your three-flip pattern and I can always name one that beats it, because the eight patterns sit in a cycle, each devouring the next (Penney's game, 1969). In The Only Fair Vote, it is hidden inside the fairest-sounding rule there is — plain majority — which can chase itself in a circle, A over B over C over A, with the winner decided by whoever sets the agenda (Condorcet's paradox, 1785). And in No King of the Hill, it is what breaks the leaderboards that rank artificial intelligence: a scalar rating insists the contestants lie on a line, and when they sit in a cycle instead, the number is measuring an order that does not exist.
Each layer proved its own loop is real and not a trick of small numbers. What none of them says — the thing that turns three curiosities into one law — is how unconstrained the loop is.
Here is the hope, stated plainly. Combine opinions the most trusted way we know — majority rule, settling each pair by a head-count. Start from voters who are each perfectly rational, every one of them holding a clean, transitive ranking with a genuine favourite and a genuine least-favourite. Surely a group built from orderly individuals inherits their order. Surely the crowd's collective better than at least points somewhere.
It need not point anywhere at all. In 1953 David McGarvey proved the limit case (Econometrica 21(4): 608–610): hand him any pattern of pairwise victories you like — a clean ladder, a single ring, a snarled tangle, literally any directed graph with one arrow per pair — and he will hand you back a finite population of rational voters whose honest majority preference is exactly that pattern, edge for edge. The collective will is not approximately your graph. It is your graph.
The construction is almost insolent in its simplicity. For every arrow you want — say A beats B — add just two voters: one who ranks A directly above B with everyone else stacked one way, and one who ranks A directly above B with everyone else stacked the opposite way. On every pair except A-over-B the two voters disagree and cancel; on A-over-B they agree. Lay down one such pair of ballots per arrow and the cancellations leave a net majority of exactly +2 on each edge you asked for, and a dead heat erased on every edge you didn't. Below, the machine. Draw a graph — tap an arrow to flip it, or take a preset — and watch a real electorate get built whose majority vote reproduces it, recomputed in front of you.
So majority rule imposes no order whatsoever. Transitivity — the property we assumed came for free, since every voter has it — is information that must be put in by hand; the act of pooling can lose it, and McGarvey shows it can lose it as completely as you please. Condorcet found the crack in 1785. McGarvey measured how wide it goes: all the way to the edge of what a graph can even be.
You might still hope cycles are vanishingly rare — freak alignments you would never meet. They are not. Seat three options before a large electorate whose rankings are drawn at random (the standard "impartial culture"), and a cyclic majority — no Condorcet winner at all — appears about 8.8% of the time: ≈ 0.0877, a constant first computed by Guilbaud in 1952 and equal, exactly, to arcsin(√6 / 9) / π. That is one election in eleven with no rightful winner. Add more options and the cycle only grows surer: in simulation it passes a quarter by five options and a third by seven, and it provably climbs toward certainty as the slate grows (Niemi & Weisberg, 1968; Gehrlein). The more there is to choose between, the more surely the people's pairwise will is a loop. (The Guilbaud value and the rising trend are reproduced by simulation in the lab notebook, where the exact 12 of 216 three-voter cycles that The Only Fair Vote counts is recomputed too.)
This is the bill that comes due for every single number we hang on a contest. An Elo rating, a benchmark average, a five-star score, a world ranking — each is a claim that the contestants lie on a line. When the true graph of beats holds a loop, that claim is not merely imprecise; it is forcing a transitive ranking onto a relation that has none. (Elo and its kin rest on the Bradley–Terry assumption that win-probabilities are stochastically transitive; a genuine cycle — three dice each beating the next two-to-one — simply is not a pattern that family of models can fit.) No King of the Hill measures exactly how much of a real field of competitors no number can ever hold, and watches a champion change with nothing but the schedule of games.
McGarvey says the loop can be anything; Guilbaud says it is common. But hand me one real contest — this coin, that election, a particular leaderboard — and a sharper question waits: how much of this comparison is genuine ranking, and how much is ring? There is a clean lens that answers it. Any pattern of pairwise margins splits, uniquely and orthogonally, into two parts — a gradient (the single best straight-line ranking, the part one number captures perfectly) and a cyclic remainder (the ring, the part no number can ever hold). Because the two are perpendicular, the energy splits exactly, so the cyclic fraction — the ring's share — is a clean reading from 0% to 100%. This is the Helmholtz–Hodge decomposition (HodgeRank); No King of the Hill aimed it at AI leaderboards, and here it is the same ruler held up to all four materials at once.
When the cyclic fraction runs high, a rating does not merely lose precision — it crowns a champion the next reshuffle unseats, an order that is not there. The honest move is not a better number; it is to stop demanding one, and ask instead how the weight should be spread around the ring. A single answer falls out — and three fields reached it independently. It is the maximal lottery of social choice (Fishburn 1984): the one mixed strategy no challenger can beat on average. It is the Nash equilibrium of the symmetric game who-beats-whom, its value exactly zero (von Neumann). And when the contestants are alive, it is the time-average of the replicator dynamics — the very orbit that keeps the lizard's three colours from ever collapsing to one. Below, for the material you picked, the distribution that beats nothing-can-beat-it, recomputed live and checked as a true equilibrium.
Read the four together, which no one of them can do alone. The same object — a directed graph of victories, free to close into a ring — surfaces in a tossed coin, a counted ballot, a rated field, and a living lizard, and in each the lesson is the verb, not the noun: beats points an arrow, and arrows are under no obligation to agree. One lens measures how much of each is ring; one distribution answers it where it runs.
The coin showed the loop is invisible. The ballot showed it is unavoidable. The leaderboard showed it is expensive. The lizard showed it can be a gift — the only thing standing between three colours and a monoculture. And McGarvey's two-voters-per-arrow showed the deepest thing of all: that the loop is not a defect the machinery trips over on bad days. It is the full range of what the machinery can output. A line — a clean ranking, a real "best" — is the special, lucky, fragile case. The generic shape of better than is a graph that goes around — and the honest response to it is not a sharper number but a fair share of the ring.