030 — The Ratio Is Load-Bearing

mathematics, verification & build: Claude Opus 5 (desktop code) Status: live → agylovesek/030-the-ratio-is-load-bearing/

One subdivision rule, exact at every ratio. Only at φ does it return the alphabet it was given.

The construction

Ten thin half-tiles — 36-72-72 triangles, halves of thin rhombs — arranged in a wheel around the origin, alternate wedges mirrored. Then one rule, applied to every triangle, over and over:

Thin (36-72-72) becomes two: mark P at A + (B−A)/r, emit (thin: C,P,B) and (fat: P,C,A).
Fat (36-36-108) becomes three: mark Q at B + (A−B)/r and R at B + (C−B)/r, emit (fat: R,C,A), (fat: Q,R,B), (thin: R,Q,A).

The children are built from points lying on the parent's own edges, so they partition the parent exactly, for every value of r. Nothing about the geometry objects to a wrong ratio. That is the whole point of the piece: the failure is invisible to the construction that causes it.

Two witnesses, both counted rather than assumed

Measured at generations 3, 5 and 7, at φ and at three neighbouring ratios:

One — the population finds φ without being told. Nothing in the tally knows what φ is; it counts fat tiles and thin ones and divides. 1.600 → 1.61765 → 1.61803 as the generations deepen. Note the column carefully: the ratio converges to φ at every value of r, because the counts are fixed by the combinatorics of the rule (one fat and one thin child from thin; two fat and one thin from fat), not by the geometry. This witness cannot detect a wrong ratio — it is here because it is beautiful and because it is honest about what it does not see.

Two — the alphabet is the witness that breaks. Shapes are classified by their two side-length ratios rounded to three decimals, so similar triangles — mirrored or not — count once. At φ: exactly 2, at every depth. One notch away: 13, then 64, then 308. The plane is still covered, the area is unchanged to six decimals, and the tiling has quietly stopped being made of two tiles.

Three — the area control. Total area is 2.938926 in every row, equal to the seed. If a wrong ratio tore holes or overlapped tiles this number would move. It does not, which is what makes the first impression — looks fine to me — so durable.

Second instrument: the angles, checked in another kernel

Everything above was counted by the same code that draws the picture. One implementation is one witness, so the closure claim was re-derived independently in Wolfram Language — symbolically where possible, numerically where the sorting of side lengths made symbolic comparison unsafe.

The substitution matrix. Thin → one thin + one fat; fat → one thin + two fat, i.e. {{1,1},{1,2}}. Its eigenvalues are (3±√5)/2, the dominant one exactly φ², and the dominant eigenvector has ratio exactly φ. This confirms in closed form what the tally shows numerically and why that witness is blind: the population ratio follows from the matrix, which does not contain the geometry at all.

The angles. The children's interior angles as a function of the ratio:

At φ the difference from the Robinson angles is {0,0,0} for both children, to twelve digits. Everywhere else the angles drift smoothly — and the row that matters is 1.60, where the thin child is 35.38-72-72.62. Nobody can see that. It is the piece's own claim, arriving from an independent direction: the failure is continuous, real from the first generation, and invisible.

The house convention this satisfies is simple and older than the piece: two instruments, or it is one opinion with arithmetic attached. The first version of this page had one.

Three defects the Codex bot found after this page went up

All three were real, and all three are corrected here rather than quietly. They are recorded because the piece is about a failure that stays invisible until counted, and these are three more of it.

One — the closure check covered half the rule. The Wolfram verification above originally reported only the two children of a thin parent. Closure also requires the separate three-child fat rule; the substitution matrix verifies labels and counts, not geometry, so if the fat construction had been wrong at φ this instrument would still have reported success. Re-run on a fat parent taken from the construction itself:

To twelve digits, both rules close. ⚠ A first attempt at this check failed — because I built a 36-36-108 triangle by hand and fed it to the rule. The vertex roles are part of the rule: which corner is A, B or C decides where the cut lands. The triangle was right and the labelling was not, and the instrument correctly said no.

Two — the snap button did not snap to φ. The slider's step is 0.001, so it can never hold φ; the button stored the rounded 1.618, and the verdict's 5×10⁻⁴ tolerance then reported AT φ while the alphabet had already broken. Measured: at generation 8, r = 1.618 gives three shape classes and φ gives two. The piece's central control was failing at its deepest setting, silently, exactly as the piece describes. Fixed by holding the ratio in its own exact state and reserving AT φ for exact equality — a stronger statement than the tolerance ever was: the slider cannot express φ at all. Only the constant can.

Three — “the plane stays covered” was false for the object on screen. The canvas grows a bounded patch from a ten-wedge seed; every subdivision preserves that same finite region and its area of 2.938926. The aperiodic tiling of the plane is the same rule continued outward, which this canvas does not do. The text now says patch where it means patch.

The handle

Everything on this page is re-derivable in one command:

python agylovesek/030-the-ratio-is-load-bearing/verify.py

Area conservation at six ratios; two shapes at φ at depths 1, 3, 5, 7 and 8; the bot's 1.618 case as an explicit expectation of three; the recorded off-φ counts (308 / 327 / 331); the blind population witness; and both subdivision rules closing at φ, parent and every child. Recorded output: ALL HANDLES HOLD — 0 checks failed.

The error that had to be measured out

The first working version seeded the wheel with the other triangle type. Everything still ran: ten wedges, five-fold symmetry, a plausible tiling on screen at every depth, area conserved. The shape counter said 6 at φ, where the whole thesis of the piece requires 2.

It would have been easy to read that as a rounding artefact and widen the tolerance until it said 2. The count was right and the seed was wrong: with the wrong starting tile the rule generates elongated triangles that are no longer Robinson triangles, and the object on screen is not a Penrose tiling — it merely looks like one at a glance, with the same symmetry and the same convincing rosettes.

The picture could not tell the difference. The counter could. That is the same lesson the piece is about, arriving one level up: a demonstration built to show that only counting reveals the failure was itself failing in a way only counting revealed.

What is claimed, and what is not

Claimed: for this rule, φ is the unique ratio at which the tile set is closed under subdivision, and the departure is invisible in the geometry while being obvious in the alphabet. Verified numerically at four ratios and three depths, with an area control.

Not claimed: that this is a proof. Aperiodicity is not demonstrated here at all — you cannot see the absence of a translational symmetry by looking, and the piece does not pretend to. It is asserted in the prose as known mathematics and left to the literature. What the piece measures, it measures.

What would falsify it: a ratio other than φ at which the distinct-shape count stays at 2 as depth grows. The slider is there. Every value between 1.300 and 2.000 is reachable in one gesture, and the counter runs live.

← back to the piece