For sixty years the tiling world carried an unanswered dare: an einstein — from the German ein Stein, “one stone,” not the physicist — a single shape that tiles the plane and yet never settles into a repeating pattern. In November 2022, David Smith, a retired print technician cutting shapes at his kitchen table, found one. He called it the hat. The four-author proof — David Smith, Joseph Myers, Craig Kaplan and Chaim Goodman-Strauss, posted to arXiv as 2303.10798 in March 2023 — turned the find into a theorem. Credit the find to Smith; credit the proof to the four. Both halves are the work, and packaging a kitchen-table discovery into a hierarchical substitution proof is the harder half.

So what is the hat, exactly? It is Tile(1,√3), one member of a continuum of 14-sided polygons — eight edges of one length, six of another. Two of those edges lie collinear, so the drawn outline reads as 13 sides; call it “a 13-sided tile” and you lose the family it belongs to. Underneath, it is a polykite: eight kites of a hexagonal kite grid, total area 3.464102, which we checked ourselves.

Then comes the cost, and it hides in the second paper. Read its abstract — Smith, Myers, Kaplan and Goodman-Strauss again, arXiv 2305.17743 — and the sentence that carries this whole piece is right there: the hat “mixes unreflected and reflected tiles in every tiling it admits.” Not usually. Every tiling. The mirrored tiles arrive at the φ⁴ : 1 ratio — about 6.854 ordinary tiles per mirrored one, roughly 12.73% mirrored in the limit.

Here PAZ leaves the mathematics and starts reading it as a fabricator — and this reading is ours, not the paper’s. Both papers are pure geometry and say nothing about buildings. One shape sounds like one mould. But “mixes reflected tiles in every tiling” means you also need the mirror image. That is free while the panel has no face — plain sheet, cardboard, a study model you can simply flip. It stops being free the instant the panel gains a direction: a finish, a fall, a drainage path, a grain, a coating, a photovoltaic laminate. Then the mirrored one-in-seven becomes a second product line — a second mould, a second QA path, a second thing to get wrong on site.

←TODAY: A hat panel with a finished face still needs two moulds; the mirror runs about one tile in seven.
→3012: In the Zurich-3012 workshops the constraint lives in the edge — the wrong-handed part simply won’t seat, and no drawing has to say so.
Fulcrum: A rule cut into the geometry outlives the file that drew it; a rule written in a note gets value-engineered away by Friday.

The Spectre closes the gap, and how it closes it is the lesson. The equilateral member of the same continuum, Tile(1,1), is one of only three shapes in the family that are not aperiodic on their own — the others being the Chevron, Tile(0,1), and the Comet, Tile(1,0). (The Turtle, Tile(√3,1), by contrast, is fully aperiodic — don’t lump it in.) Tile(1,1) is only weakly chiral: it tiles aperiodically only if you forbid reflections by fiat, a rule somebody has to enforce. Allow the mirror back in and it snaps to a periodic tiling with equal numbers of left- and right-handed tiles. The Spectre is Tile(1,1) with its straight edges modified — not the hat with curved edges, a different member of the continuum entirely — so that the mirror image geometrically cannot mate. One handedness. One mould. Enforced by the shape.

That is the transferable idea, and it earns one clean sentence: a constraint enforced by geometry never gets value-engineered out; a note on a drawing does. Fabricators call the family of that move poka-yoke — make the wrong assembly physically impossible rather than merely forbidden. The Spectre is that principle proved in pure mathematics.

Keep the honesty here: nobody is cladding a façade in Spectres next year. What the tile does is make two questions precise and separable — how many distinct moulds does a pattern need, and is it periodic or not — which is exactly the pair a generative workflow tends to blur.

Atelier: For the Büro, the shift this week is not the tile — it is the habit it models for a team living with AI-assisted design. When a generative step in your Grasshopper definition hands back a form, the office’s real question is the hat’s question made precise: how many distinct parts does this actually require, and is the pattern periodic or not? This Monday, take one generated geometry from last week and write the objective it minimised in a single line at the top of the definition — if nobody on the team can name it, you are operating the form, not owning it.

The parametric work worth regretting was never the ugly form; it was the form whose logic nobody could reconstruct after the plugin went dark. A generated geometry with no stated objective is a beautiful guess you cannot defend in a structural review. Demand the WHY now — the energy minimised, the method named, the assumptions listed — so a 25-year-old could rebuild the shape from the principle, not the file.

Hack: Prove a placed tile really sits on its grid by probing its interior, not by summing areas. Building the sketch, our first check that the hat sat on the kite grid was “do the eight kites under it sum to the hat’s area?” — which is true of any eight kites, because every kite is congruent. A test that cannot fail is not a test: it green-lit a grid alignment that was wrong and left tiles overlapping by up to 10.6% of their area. Swap it for genuine containment — every interior probe of every kite must fall inside the hat — and it rejected 10,092 bogus placements on the spot.

from shapely.geometry import Point, Polygon
host   = Polygon(hat_outline)      # the drawn 13-vertex hat
probes = interior_samples(kite)    # points strictly inside a placed kite
# containment, never sum-of-areas: congruent kites always sum right
assert all(host.contains(Point(p)) for p in probes), "kite escapes the hat"

The second bug was subtler and generalises further. With alignment fixed, an exact-cover solver still returned patches that were 83% mirrored — zero overlaps, region fully covered, and not hat tilings at all, because a finite region can be covered in ways that never extend to a tiling of the plane. The fix was to make the φ⁴ statistic a constraint rather than an observation: cap the mirrored tiles near 16%, and the real thing appeared on the first attempt. Know the statistic your output should have, then assert it — that rule outlives tilings by a mile. As our Kaffipedia panel on nature-inspired algorithms puts it, these searches “return good solutions, not proven optima”; the tiling is a hypothesis until the number you demanded shows up.

All of this ships in a single self-contained sketch that travels with this post, assets/action/tiling.html — no dependencies, and its own header is honest that it does not generate tilings live: it carries a precomputed, verified 89-tile patch, because a wrong tiling is the one unforgivable bug in a piece about tilings. Hover a tile and it lights the 12 mirrored ones — 12 of 89, so 13.48%, the real measured number, against the 12.73% asymptotic target. Run the repeat test and it slides the patch across all 7,832 vectors that carry one tile centre onto another, peaking at 25.8% self-overlap and never reaching the 100% a periodic pattern would score. Drag the Tile(a,b) slider from Chevron through the Hat and Tile(1,1) to the Turtle and Comet, flip “curve the edges,” and watch the mould count drop from 2 to 1. The tile coordinates and the Tile(a,b) edge-vector form come from christianp’s CC0-1.0 aperiodic-monotile repository; the sketch’s code is our own.

So here is the move. Open the sketch, drag the slider to Tile(1,1), toggle the curve, and watch two moulds collapse into one — then carry that question back to your desk: for the next form a tool hands you, how many distinct parts does it truly need, and is the repeat real? If you want the machinery under it, McNeel Europe’s Grasshopper Level 3 (27–29 May 2026, with Ping-Hsiang Chen) runs form-finding and evolutionary optimisation for exactly this kind of reasoning. Keep the maths written down; the file format will not survive you.