A fancy lookup table

Strip away the vocabulary — the breathing cycles, the rotors, the certification wall — and here is what our reasoning machine really is: a lookup table with fuzzy matching and grouping.

That sounds deflationary. We mean it as the design’s proudest honesty.

The machine maintains atlases: maps of the operation-shapes it knows, kept in two spaces at once — the language space (the many ways English expresses an operation) and the math space (what the operation actually does). Each entry is a centroid: the average location of a known kind, the center of gravity of every example of that kind the machine has seen. Reading a new problem means fuzzily matching its silhouette against the atlas, grouping it with its kind, and retrieving the exact machinery for that kind.

Keeping an atlas honest takes bookkeeping discipline. Centroids are maintained with Welford’s algorithm — the numerically careful way to update a running mean and variance as examples stream in, one at a time, without ever holding the whole history in memory. And the atlas cannot simply be kept forever: we are now on the seventh generation of centroids, one per era of the trained head, because each new generation of weights rotates the internal coordinate system. The rotation is nearly pure — aligned generations agree at cosine 0.988 — but “nearly” is doing real work in that sentence, and so every generation the atlas is re-anchored from scratch. Old maps are never trusted in new coordinates. That rule has caught more subtle bugs than almost any other in the project.

Why isn’t this just memorization with extra steps? Because of what the table’s keys are. A memorizing system keys on surfaces — sentences, phrasings, numbers — and shatters the moment a costume changes. Our table keys on silhouettes: the structural shapes that survive the 512-dimension waist after the costume is destroyed. Ten thousand differently-dressed problems collapse onto a few hundred keys. The table stays small; the coverage stays wide; and a brand-new costume on a known dance looks up correctly on the first try.

There is a respectable philosophical position that all cognition is sophisticated retrieval — that expertise is less like derivation and more like a chess master recognizing fifty thousand positions. We take no side in that debate for humans. For machines we’ll say it plainly: retrieval with the right keys, over an honestly-maintained atlas, backed by exact verification of whatever gets retrieved, is not a lesser form of reasoning. It is the form of reasoning you can audit — every lookup names the kind it matched, every kind names the examples that built it, and every answer either survives the solver or never leaves the building.

A fancy lookup table — with its keys chosen so well that looking up becomes understanding.