title: Two loops, one shape date: 2026-09-14
The machine scores about 0.25 on wild prose and about 0.98 on synthetic problems of the same algebra. That gap is the whole project right now, and this week we learned where it lives to the field. Decoding every breath’s state and scoring each field, wild gets presence, type and result pointers at 0.88, 0.84 and 0.78, and fails on two things: argument pointers at 0.39 and digits at 0.32. The machine knows a relation is there and roughly what kind. It cannot say which quantities the relation binds, and it cannot read a value off ordinary sentences. The math is not the blocker. Reading is.
Four organs touched the token roads, each with bars pinned before the read, and each left wild where it was, inside one standard error of the measurement. A token convolution learned to be the identity. A steering wheel that melted contradictory slots and forbade their refuted bindings walked parses to consistency, and consistency is not correctness: a wrong parse that satisfies itself draws no contradiction. A structural claim mask, tokens marked as taken once a slot binds them, gave half a point and stayed flat across its whole sweep. And an aligner that anchored each given’s value to its digits in the text leaned negative on both seeds, for a reason worth stating: the variable is the quantity, “the base coat’s drying time,” and the numeral is its value. Anchoring the variable to the numeral taught the wrong thing.
The common lesson is that the wild register lacks something no organ manufactured: knowledge of what ordinary phrases mean as quantities and operations. The books supply it by hand, one annotated row at a time, and the books scale. Nothing automatic reproduced it.
The project’s founding rule is that abstraction may live in annotation and recognition, never in verification. The solving jaw is a general constraint core with zero domain code; the answer key gates everything; nothing on that side learns. The construction jaw proposes, and that is where knowledge about the world is permitted.
Bryce’s proposal, which I now think is the right one, follows that rule to its conclusion. The language side should have a curated table of what prose means: “a dozen” is 12, “twice as many as” is multiply by 2, “half of” is a division, “each” and “per” are a multiplier over a count, “more than” is add, “fewer than” is subtract. Recognition knowledge, symbolic, written once, zero parameters. It is the language side’s macro library: a floor of the ladder that is expanded before the neural organs see the sentence, exactly as the math side’s macros are expanded before the solver sees the graph.
And it lands on the two fields that fail. Values arrive in prose as words and multipliers, not numerals, and the head has only ever seen numerals; a table turns “twelve” into a value mention with a span. Operators hide inside “twice as many as the boys”; a table turns that into a typed anchor with its operand’s position. Those are the footholds the grounding roads have been missing, and no trunk layer gives them.
Look at what the math side has, and the shape is three organs. An atlas: kinds, centroids, macros, the knowledge of what problems look like. A certifier: the constraint solver, which disposes of what the neural organs propose. And dynamic masks: the structure the machine routes messages along, rebuilt every breath from what it has committed.
The language side can have the same three, in its own medium. Its atlas is the lexicon. Its certifier is the fingerpost: the machine already certifies parses by unanimity across five permuted views of the sentence, on the principle that a correct reading is the same from every angle and wrong readings each differ in their own way, and a reading that holds across views is the language-side analog of a parse that satisfies the solver. Its masks are claims: a token taken by one slot is marked so other slots read around it, and a slot the certifier melts releases its tokens. Two loops, one shape, each proposing neurally and disposing symbolically in its own medium, with the waist between them as the handoff.
This also settles a question we spent a day on. The frozen trunk stays four layers of Llama. Looping frozen layers is a refuted cell in our own ledger: a layer trained to consume its predecessor blurs when it consumes itself. And deeper layers, which were the next candidate, are the most expensive parameters in the system. If the language loop’s knowledge comes from a table and its certificate from the fingerpost, the trunk does not need to grow.
A lexicon must be a thin anchor layer that feeds the neural roads, never a grammar. The moment it tries to parse sentences we have rebuilt the 1990s, and the trunk exists so we don’t have to.
It covers the head of the distribution, and prose has a tail. So the first step is a census, not curation: how many of the pen rows’ givens are numerals, how many are number words, how many are multipliers, and how relations are phrased. That sizes the table before anyone writes it.
And coverage is not correctness. Each entry is a claim about meaning. The fingerpost should certify entries the way it certifies parses: an entry that produces view-invariant readings on the rows it fires on stays, and one that doesn’t is struck.
Census first, on the CPU, no training. Then the aligner redone into the value channel, with numerals and number words, two seeds against the existing controls, bars in standard-error units. Then operators and multipliers as the lexicon’s second rung. Then the claims as the language loop’s masks, protecting anchors that finally exist. The depth probe waits behind all of it, because if the lexicon moves wild, depth was never the question.
A convolution bets that neighbors mean something. A blur in the loss bets the middle can be graded on its own work. A lexicon bets that the machine has been asked to rediscover, from 0.25 accuracy, what every reader already knows about a dozen. That bet is the one on the board this week.