Rotations: same shape, another angleWhy two completely different systems can hold the same meaning — and how you would check

Some of the most surprising claims in this library sound like this: two language models trained by different companies on different data hold their words in the same shape, just rotated. Or: a brain trajectory and a text trajectory line up after one rotation. To feel what those claims mean — and what would make them true or false — you only need two ideas: what a rotation preserves, and how the dot product measures agreement. Both fit on this page, and you can drag both.

1. What a rotation is: everything moves, nothing changes

Take a little constellation of points — stars, or cities on a map, or (hold the thought) words. Now turn the whole thing, rigidly, around its center. Every point moves. And yet, in the sense that matters, nothing has changed: every distance between every pair of points is exactly what it was. Every angle. Every neighborhood. Who is close to whom, who is far from whom — the entire web of relationships — is untouched. A rotation is a change of where that preserves all of how things relate.

That is the whole definition. The heavy notation you may meet elsewhere on this site — "a rotation R ∈ SO(D)" — says only this, in D dimensions instead of two: a rigid turn that preserves all distances and does not mirror-flip.

Turn the constellation. The readout tracks three of its internal distances: they do not budge, at any angle. Position is relative to you; shape belongs to the constellation itself. Shape is what rotation cannot touch.

2. The dot product: a number for agreement

Now two arrows from a common origin. How much do they agree — point the same way? The natural measure: let one arrow cast a shadow onto the other's direction and ask how long that shadow is. That shadow-length (times the other arrow's length) is the dot product. Arrows aligned: big and positive. At right angles: zero — no agreement, no opposition, just unrelated. Opposite: maximally negative.

If both arrows have length one, the dot product is exactly cos(angle between them) — the same cosine-as-agreement you may have met in the phase primer or Foundations 1, where the "arrows" are phases on a circle. One idea, everywhere: cosine measures agreement. When the essays score how well two things line up — two oscillators, two attention heads, two whole minds — they are composing this one little measure at scale.

Swing arrow B and watch its shadow on A. The dot product is the shadow length: full agreement at 0°, zero at 90°, full opposition at 180°. Sound on: loudness is the agreement — the same mapping as the phase figures, so your ear learns one language for the whole library.

3. "One rotation apart": aligning two star maps

Now put the two ideas together. Suppose two systems each hold the same set of concepts as points — the same shape — but each at its own arbitrary angle (why would two separately-built systems agree on an orientation?). Looking at raw positions, they seem to disagree about everything. But if the shapes truly match, there exists one single rotation that lays every point of one onto the corresponding point of the other, all at once. Find it, and the correspondence snaps into view.

Try it by hand:

Map A (blue) and map B (orange) hold the same seven concepts; B was turned by a hidden angle and lightly jittered. Slide until the maps snap together — the match score is the average agreement between corresponding points. One number, one knob, and suddenly two "different" systems are revealed to hold one shape. match = —
What's real here This figure is a cartoon built from seven made-up points in two dimensions, so you can feel the logic in your hands. The real measurement it is a cartoon of lives in The Physics of Mind §15: two independently trained language embedders, hundreds of dimensions, and a single rotation that aligns them at 99.3% held-out top-1 accuracy. The claim "meaning has one shape" is exactly the claim that this game is winnable there — not just here.

Where this goes

With rotation-preserves-shape and dot-product-measures-agreement in hand, you can now read three of the site's bigger moves: the shared manifold of meaning (two models, one rotation apart), the brain↔text alignment (the same trick across substrates — there the rotation is called the Canonical Resonance Transform), and — one primer further — holonomy: what it means when carrying something around a loop brings it back rotated, which is where this library locates memory and self-observation. That primer is coming; until then, the gentle on-ramp is §10 itself.