2026-09-05
Our machine keeps two maps of what it knows. The language atlas charts how each kind of mathematical operation appears in English — its silhouettes, its costumes. The math atlas charts how each kind behaves during thinking — and recently that chart grew seven pages per kind, one for each cycle of deliberation, because a kind’s location early in thought (loose, exploring) is not its location late (tight, committed). A kind is not a dot; it is a trajectory.
That created an asymmetry that bothered us. The math atlas got seven pages; the language atlas kept one. And the reason seemed structural: our frozen language model reads the problem exactly once — one forward pass, one representation, then it never speaks again. You cannot have seven snapshots of a thing that only happens once. Can you?
You can. Because the representation is not the reading.
The frozen model’s output is static — a rack of costumes that never changes. But every cycle of deliberation, the machine’s attention re-reads that static representation with different weights, lingering on different words as the parse commits. The text doesn’t change; which parts of it matter changes, breath by breath. Pool the frozen states weighted by each cycle’s attention, and you get a sequence of language-states that evolves — for free, from a model that ran once. Same rack, different pose each breath. The dancer doesn’t re-sew the dress seven times; she moves through it.
So the language atlas gets its seven pages after all: not “how the text embeds” but “which parts of the text matter at each stage of thinking.” Early pages should show broad, surveying reads; late pages should narrow onto the few spans that carry the answer — the same nested-doll contraction we measure on the math side, now visible on the language side.
The atlases share an index: one kind, two coordinate systems — a road map and a topographic map of the same country. And the transport between them is the act of reading itself: parsing carries a point from the language chart to the math chart. Making that transport explicit buys something new — when a problem arrives, its language-silhouette can pre-fetch the math trajectory its kind usually follows, handing the machine’s attention controller a prior before the first commitment: this smells like a ratio problem, and here is the road ratio problems usually take.
The best part comes from holding both seven-page charts at once. Each kind becomes a paired trajectory: what-I’m-reading and what-I’ve-committed, side by side, cycle by cycle. And pairs can be compared: does attention narrow onto the answer-bearing words before the commitments form — reading driving mathematics — or after — mathematics driving re-reading? That lead/lag needle, measurable per kind and per cycle, is the sharpest instrument we’ve ever designed for observing the alternation this whole project is built on. The telegraph signal we found years ago in attention traces was this rhythm heard through a wall. The paired atlas is a stethoscope pressed directly against it.
An asymmetry that bugged us turned out to be a measurement we’d been pooling away. The model runs once. The reading never stops moving.