Crossing the bridge: from neural to symbolic

Our reasoning machine lives on both sides of the deepest divide in computing, and its whole job is to carry meaning across the bridge between them. This post is that bridge, drawn slowly, with every term defined.

The two sides

The neural side is continuous and non-deterministic. - Neural: computation done by a network of learned weights — millions of numbers tuned by training, no one of which means anything alone. - Continuous: everything is a shade. A neuron’s activation is 0.73, a similarity is 0.988; nothing is simply yes or no, and every representation lives in a smooth space where “between” is always defined. - Non-deterministic (in practice): outputs are graded beliefs, sensitive to training randomness and input phrasing. Ask twice, in two costumes, and you may get two answers — which is exactly why our certification re-asks in five costumes and demands agreement.

The symbolic side is discrete and deterministic. - Symbolic: computation over explicit tokens with exact rules — variables, operations, constraints. Everything is legible: you can print the state and read it. - Discrete: no shades. A variable equals 7 or it doesn’t. A constraint is satisfied or violated. There is no 0.73 of a fact. - Deterministic: the same input always yields the same output. Our solver run twice gives byte-identical answers, forever. Its refusals are as reproducible as its solutions.

Neither side can do the other’s job. Wild language is a continuous phenomenon — costumes, shades, ambiguity — and only the neural side can grip it. Truth is a discrete phenomenon — right or wrong — and only the symbolic side can guarantee it. The law of the house: neural proposes, symbolic disposes. The bridge carries proposals one way, and verdicts back.

The bridge itself

The crossing happens in stages, each destroying a little more continuity. A sentence enters as 2048 continuous numbers per word; the waist squeezes each to 512, destroying phrasing and keeping structure; cycles of deliberation sharpen graded beliefs into committed choices; and at the far end the parse snaps to discrete form — a typed factor graph, all integers and named operations, which the deterministic solver either solves exactly or refuses legibly. Continuous in, discrete out, and the moment of snapping is the single most important event in the machine.

The two atlases

Recognition on the neural side is guided by atlases — maps of the operation-kinds the machine knows, one for each side of the bridge:

• The language atlas maps how English expresses operations: the many surface silhouettes of “multiplication” or “a total.” Today it chiefly guards the door — an out-of-distribution check asking “is this sentence from territory we know?” before the machine is allowed to answer.
• The math-operation atlas maps what operations are: each known kind summarized as a centroid — the running average location of every example of that kind in the machine’s internal space. Centroids are maintained with Welford’s algorithm, the numerically careful way to update a mean and variance one example at a time without storing the history. We are on the seventh generation of these centroids — one per era of the trained head, rebuilt every era because new training rotates the internal coordinate system (nearly purely: aligned generations agree at cosine 0.988), and an old map must never be trusted in new coordinates.

An open question, honestly stated

The two atlases are not treated equally. The math-operation atlas is consulted, re-anchored, audited every generation. The language atlas mostly stands at the door. We suspect an imbalance there — that the reading itself could lean on the language atlas the way navigation leans on a chart, not just as a border checkpoint. It is on our books as a registered question, which in this project means: it gets a pinned prediction and a measurement, and the ledger records the answer either way.

Two sides, one bridge, two maps — and every crossing audited.