Tukey’s fence moved into hyperbolic space. Each point is measured by the mass it reaches, the volume of the ball about a robust centre out to it. Tukey’s own fences, Q3 + 1.5 IQR and Q3 + 3 IQR, are then applied to that mass. On a cloud spread the way a hierarchy spreads, the fence lands on the doubling sphere of the upper quartile. In the hyperbolic plane that sphere sits less than ln 2 = 0.693… beyond the quartile, and approaches that gap as the cloud grows. That is the 0.7. In the plane every verdict is decided exactly; in H³ every verdict carries a certificate. The tracer explainer follows one run stage by stage and proves the lemmas, and the fence in 3D shows H³ as nested shells.
Seen from the centre: each point at its hyperbolic distance, in its direction. The circles are the fence’s spheres. Click a point to read it; in H³ drag to turn.
Top: the mass coordinate, in units of the upper quartile’s mass, where the Claude fence acts. Bottom: hyperbolic distance, where distance-Tukey acts. The quartiles pick the same points in both (CF1); only the fences differ.
In the plane a verdict is two comparisons of rationals: 2b > 5b₃ − 3b₁ and b > 4b₃ − 3b₁. In space the mass sinh 2d − 2d is evaluated in fixed point with a proved error bound, doubling the bits until the sign is certain. A point inside the upper quartile needs no arithmetic at all: its order decides it.
No four points of hyperbolic space have a four-point defect above ln 2. Corrupt some of the cloud’s distances: any quadruple that now beats ln 2 proves one of its six distances wrong, and the pairs in the most such quadruples are the suspects.
On load: the Python reference’s fence trace, replayed through this page bit for bit. On request: the nine lemmas checked live, each with a control that must fail.