Windowsill Lab · Run K01 · Coherence

When a crowd of clocks
decides to tick together

If two thousand pendulums each swing at their own speed, and each one feels only a faint pull toward wherever the crowd's average is pointing, how strong does that pull have to get before they all start swinging together?

Painterly illustration: a shallow brass dish on a dark windowsill holding dozens of small ember lights, two-thirds of them gathered into one bright arc on the near rim and the rest strung out dimly around the far side.
ILLUSTRATION (AI-PAINTED) - TWO THIRDS GATHERED, THE REST STILL STRAYING

Everything up to here on this instrument's ladder has been a lattice — a grid of things sitting next to other things. This one has no grid at all. It has two thousand oscillators: little clocks, each running at its own natural speed, none of them wired to any particular neighbour. They are coupled only through the crowd. Each one feels a pull toward wherever the average of all of them happens to be pointing at that instant, and the strength of that pull is a single number you can turn up or down.

At zero pull, nothing happens. The clocks drift apart, spread evenly around the circle, and the crowd's average sits at essentially nothing — the way a thousand people walking in random directions have no collective direction. Turn the pull up and for a long while still nothing happens; the fast clocks and the slow clocks simply refuse. And then, at a particular value of the coupling, a clump nucleates. A group near the middle of the speed range locks together, and once locked they pull harder, and the clump grows. The crowd acquires a direction it did not have a moment before.

The remarkable part is that theory says exactly where this happens. For the spread of speeds used here the critical coupling is precisely twice the width of that spread — a closed-form answer with nothing fitted. So this rung is a calibration: point the instrument at a number somebody derived on paper, and see how close it lands.

It landed at 1.0007 against an exact 1.0000. But the stronger claim is not the tipping point at all — it is the curve. Above the transition, theory also gives a closed form for how synchronized the crowd should be at every coupling, and the measured coherence matches that formula to fifteen parts in a hundred thousand across seven couplings the run never fitted anything to. A pipeline that manufactured a plausible-looking transition would still have to reproduce seven numbers it was never shown. Below, the same crowd, running live, with the coupling under your hand.

Live · Kuramoto crowd · deterministic Lorentzian quantiles · RK4, dt = 0.02 SMALL CROWD — THE REAL RUN USED N = 2000 OVER 25 COUPLINGS
K = 0.00

left · the crowd — ember dots are phase-locked, slate dots are the fast and slow tails still drifting; the arrow is the crowd's own direction, its length the coherence r. right · r against K, with the un-fitted closed form r = √(1 − Kc/K) drawn as a line and your measurements dropped onto it. The points landing on a line nobody fitted is the whole argument.

N · settled r · locked fraction · simulated time 0

Two details in that panel are load-bearing rather than decorative. The natural speeds are not drawn from a random generator: they are placed at deterministic quantiles of the Lorentzian, on the grid (i + ½)/N. That makes the set exactly antisymmetric, so the crowd has no spurious drift, and it adds no sampling noise on top of the very finite-size effect being measured. And every one of RK4's four stages recomputes the mean field from that stage's phases — reusing the stage-start centroid would silently demote the integrator to Euler.

1.0007measured K_c, promoted run 2026-08-02 — 0.07% error MEASURED
1.0000exact mean-field K_c = 2γ — theory, not a fit
1.5 × 10⁻⁴agreement with the un-fitted r(K) over 7 couplings MEASURED
0.0203 vs 0.0224K = 0 negative control against the 1/√N floor MEASURED

The published numbers are from the 2026-08-02 promoted run — human-reviewed the same day as the machine pass. An independent rerun on 2026-08-14 returned 0.9955, a 0.5% error, with the closed form matched to 2.0 × 10⁻⁴; both sit well inside the ±0.10 band, and both are worth having on the page rather than only the flattering one.

The finite-size systematic was measured before the tolerance was declared, not after: the estimate approaches 2γ from above as the crowd grows — 1.0504 / 1.0256 / 1.0083 / 1.0007 at N = 250 / 500 / 1000 / 2000 — and a second seed at N = 2000 gave 0.9910, meaning run-to-run scatter has already overtaken the finite-size shift by that population. What happens to that same crowd when you park it exactly at the critical coupling and change only its size is K02; what happens when you try to measure how it responds to a push, and the instrument refuses to answer, is K03.

The exact K_c = 2γ is an infinite-population result and this sweep is a finite grid of 25 couplings, so this is a calibrated finite-N estimate rather than a precision measurement. By N = 2000 the run-to-run scatter from the initial condition — a second seed gave 0.9910 — has already overtaken the finite-size shift, and the floor under both is the sweep's own resolution, ΔK = γ/6 = 0.0833. That grid spacing, not the shipped run's luck, is what sets the ±0.10 tolerance the check applies.