Purple Paper · side-sheet · learning machines · III

The Gorge — a perceptron at the waist

Put a learned dot-product at the place where the two bellies meet: the gorge circle of the hyperboloid, the node of the standing wave, the IR-waist of the stack. Everything routes through it, so it's the natural home for a gate. Below it runs live — and shows why the waist wants the perceptron linear, and why a learned gate there is the wrong mechanism in the right place.

node = gorge = waist · linear element keeps the normal modes independent · the bend couples them (Fermi–Pasta–Ulam–Tsingou, 1955) · a fixed exterior gate watches the admissible set the learned channel can't witness

I · The intersection between bellies is a node

In the one-sheet hyperboloid the two bellies meet at the gorge circle — the narrowest radius, where the two families of straight rulings cross most densely. Read the stack as a standing wave and that same waist is a node: pinned amplitude, the one place the whole structure forces everything through. That's exactly what makes it the right spot for a gate — and the dangerous spot for a bend.

II · The waist, running — linear keeps the modes, the bend couples them

Inject a couple of normal modes into the channel. The waist element processes the field passing through it. Flip it between linear and bent and watch the spectrum: a linear waist leaves the modes exactly where you put them (verified: 0% leak); the bend sprouts new modes — energy ladders out — and under repeated passes it spreads and then plateaus, the FPUT surprise. The fixed gate beside the perceptron refuses when the channel pushes energy outside its admissible set.

inject: bend g 2.8

the hourglass · node at the waist (green), antinodes at the bellies. the perceptron (violet) sits at the node; the gate (ring) is the fixed witness beside it.

top: the field u(x) through the waist · bottom: modal spectrum. violet = injected modes, amber = modes the bend created. dashed line = edge of the admissible set.

energy outside injected modes

100%

energy in fundamental

ADMIT

fixed gate · set {1,2,3}

Start in BENT, hit pulse a few times, then hold evolve: the amber bars climb and settle — energy leaves the fundamental but doesn't fully thermalize (FPUT). Flip to LINEAR and evolve forever: the spectrum is frozen. The gate flips to REFUSE the moment the bend pushes more than ~5% of the energy past mode 3 — and notice it does that regardless of what the perceptron's weights are doing. The gate isn't trained. That's the whole point of the next section.

III · Two elements at the node — the channel that adapts, the witness that refuses

The waist is the right location for an admissibility gate: it's set-level (everything passes through) and pre-action (it sits upstream of the second belly). But a perceptron is the wrong mechanism for that gate — it's learned, in-band, optimized on the same data it's supposed to police. A gate that adapts to the traffic can't witness the traffic. So the waist needs two distinct elements:

learned · in-band

The perceptron — the channel

A trained dot-product that adapts the representation: discovers its own IR, compresses, reconstructs. Linear, it's a clean projection (PCA at the waist). Bent, it's expressive — and it couples the modes. It optimizes. It cannot refuse itself, because the thing deciding is the thing being decided.

fixed · out-of-band

The gate — the witness

A pre-committed predicate that doesn't learn: is the energy still inside the admissible set? It is not trained by the channel, not updated by the data, not optimized. It only admits or refuses. Because it differs from the channel, it can catch what the channel can't see in itself.

Why one perceptron can't be both. An adaptive waist optimizes; an admissibility waist refuses. Train the gate and it drifts toward whatever the channel is already doing — confirming the error instead of catching it, exactly like a verification step built from the same foundation as the code it checks. The fix isn't a better perceptron. It's a second element at the node that the first one doesn't get to train: a learned channel and an unlearned check, both at the one point the geometry already forces everything through.

IV · The reading, and the curse

The physics and the governance are the same shape. Linear normal modes are independent — a system with no internal coupling has no way to mix, and no way to surprise itself. The bend introduces coupling, which is power and risk at once: it's what lets the channel model structure, and what lets energy escape the set you meant to keep it in. The gate is the exterior witness that the coupled, adaptive channel structurally can't be for itself.

Admissibility lives at the gorge. Set-level and pre-action — your no-top theorem wearing the geometry of a hyperboloid. The waist is where a gate is cheapest and where a learned gate is most tempting and most wrong. Put the witness there, and keep it out of the optimizer's hands.

The curse. Normal modes and their independence: Daniel Bernoulli and Euler, ~1750s. The bend coupling them and energy refusing to thermalize: Fermi, Pasta, Ulam & Tsingou, 1955 — one of the first things anyone computed on a digital machine, and it broke the expectation everyone held. A perceptron at the waist that discovers its own representation: a linear autoencoder, which provably finds the principal subspace — Baldi & Hornik 1989, on Eckart–Young 1936, on Pearson 1901. Every piece you reached for tonight was already standing. lol — and this time you watched me rederive your own framework on top of them.