2026-09-02
How do you recognize a waltz? Not by the color of the dress. You watch the silhouette — the envelope of motion — and the dance identifies itself regardless of who is wearing what.
Word problems are dances in costume. “Maria has three times as many apples as Ben” and “the reservoir holds triple what the tank does” are the same dance: one quantity, multiplied, equaling another. The costume — names, objects, phrasing — is infinitely varied. The dance is one of a small number of shapes.
The recognition half of our machine is built entirely around this distinction. When we looked inside the network’s own signals, the split was measurably there: one part of the state carries the envelope — what structural move the sentence is making — and another carries the carrier — which particular words made it. The head reads a sentence the way you watch a dancer: it segments the motion into moves and classifies each move, and the costume is deliberately destroyed on the way through.
The destruction is the point, and it has an address: a 512-dimension waist. Every token’s representation — 2048 numbers wide when the frozen language model hands it over — is forced through this narrow channel before any reasoning touches it. This is the information bottleneck of Tishby’s information theory made into architecture: understanding is not the preservation of information but the selective destruction of it. Kill the phrasing, the names, the order of telling; keep the quantities, the roles, the relations. If two sentences mean the same thing, the waist should make them become the same thing.
And the waist is not a gate you pass once. Deliberation lives inside it — every cycle of the machine’s thinking reads and writes through the same narrow channel, so the pressure toward silhouette never lets up. By the time a problem reaches the solver, there is no costume left. There is only the dance, drawn in its barest lines — and exact machinery can grade a bare line.
One consequence we did not fully anticipate: silhouettes are what make honesty cheap. Our certification wall re-reads every problem with sentences permuted and surface details shuffled — five costumes on the same dance. A model that memorized costumes fails this instantly. A model that sees silhouettes doesn’t notice the change. The waist, built for understanding, turned out to be the foundation of trust.