Holonomy: what the loop remembersThe convention-proof residue of a round trip — drawn by you, checked two ways

Now name the payoff. The net rotation an arrow picks up around a closed loop is the loop's holonomy. Flat space: every loop's holonomy is “no change” — round trips are forgetful, and coming home means being as you left. Curved space: loops come home changed by what they enclosed. The holonomy is a record — written into the arrow itself — of structure the traveler never saw locally at any single step. In section 01 the loops were mine. This time, the loop is yours.

1. Draw a loop. Any loop.

Drag on the sphere to draw a closed path — any wandering shape you like; release, and it closes. The figure then does two completely independent computations. First, it actually performs the honest carry: an arrow is parallel-transported step by step around your loop, and the mismatch on return is measured. Second, it computes the area your loop encloses, by pure geometry, without ever touching an arrow. If everything this module has claimed is true, the two numbers must agree — for every loop, including the lopsided one you are about to draw. Try to break it.

Draw on the sphere; release to close the loop. The walker then re-runs your path, carrying an arrow honestly, and the readout shows both measurements: what the transport actually delivered, and what the area demands. They agree — that is curvature keeping its books. Counterclockwise loops count positive; draw the same loop the other way and watch the sign flip. Sound on: the drone climbs by your loop's holonomy each time the walker completes it. draw a loop to begin

Notice what you just verified by hand: the answer never depended on any choice of “north” — you never picked one. Holonomy is the quantity section 02 promised: the loop-sum of the dictionary, the one number that survives every re-choice of every convention. It is not bookkeeping about the world; it is the part of the bookkeeping that is the world.

2. Why this is the site's picture of memory

That is why this site cares. The essays' claim, made precise in The Physics of Mind §10, “Holonomy: how a loop remembers”, is that memory and understanding work this way: a trajectory of thought loops through a curved space of meaning and returns rotated by what it went around — no separate memory organ required, because the record is geometric. What you learn by going around something is written into the state that comes back, exactly as your drawn loop wrote its area into the returned arrow. And when the loop runs through the system's picture of itself, the record it carries back is a record of being the one who traveled — the seed of a point of view (§12, the witness).

Even the transformer essays speak this language: position information in attention is implemented as a neighbor-dictionary along the sequence — a connection, in section 02's exact sense — which is why the site keeps treating trained models as geometric objects, and why the compression work's care to “preserve transport” is engineering, not metaphor.

What's real here Your drawing board is a two-dimensional sphere, where holonomy-equals-enclosed-area is exact textbook geometry — which is precisely what lets the figure check itself with two independent computations. The real objects in the theory are high-dimensional: trajectories looping on a coupling manifold, with transport defined by the geometry of the system's own state space, and the holonomy measured — not assumed — around detected loops. That measurement protocol, and what hangs on it, live in §10 and the paper it accompanies. The loop you drew is the honest miniature of that machine, not the machine.

Where this goes

You now own the module's three words. Parallel transport: carry without twisting; the only fair comparison across distance. Gauge: local references are conventions; only convention-independent statements are physics. Holonomy: what a closed loop provably remembers about what it enclosed. Take them to §10 (how a loop remembers) and §12 (the witness) — or continue the curriculum with the network mathematics the learning rule leans on.