2026-09-07
The phrase arrived in a dream, and it is a pun before it is anything else. A gadol — often spelled godol — is Hebrew for “great,” and in yeshiva usage it names a revered sage, an authority whose reading of a text settles arguments. The Large Hadron Collider is the machine at CERN that accelerates particles to nearly the speed of light and smashes them together to see what flies out. A godol collider is the old yeshiva joke: a machine that smashes two great authorities into each other at relativistic speed and observes what fundamental disputes, sparks, and books come out of the wreckage.
Our machine has two authorities that disagree all the time. They are the rival readings of a sentence.
The neural jaw reads a word problem into a diagram — quantities, roles, the arithmetic that connects them — and at every slot of that diagram it holds a distribution of beliefs, not a single answer. For most slots the belief is sharp and we commit it. But on wild prose some slots are torn: this number is the total, or it is the rate; this relation links a and b, or a and c. The head picks the likelier one, and when it picks wrong the failure has a name in our ledger, the mis-aimed pointer. We have measured that these failures are not ignorance: the internal state of the survivors that beat every repair we have built is 99.6% decodable. The machine knows almost everything. It has aimed one pointer at the wrong slot.
The ordinary fix is to ask the head to be more confident. That is the wrong instrument, and this project has a law about why: temperature is orthogonal to truth. Confidence tells you how settled a reading is, not whether it is right.
So do not ask the reader. Smash the readings together.
Take the two rival parses of a torn slot — the two gedolim — and hand each one, together with everything else the head has committed, to the exact solver. The solver does one thing, deterministically: it propagates. Given these facts and these relations, what is forced? What flies out of the collision is one of three things. A reading may yield facts and a single forced answer: it survives. It may leave the question underdetermined: it neither survives nor dies. Or it may produce a contradiction — a variable with no possible value at all — and that reading is dead, killed not by a guess about it but by the rest of the problem it could not live with.
That is disambiguation without an oracle. Neural proposes both readings; symbolic disposes of the one that cannot be true. It is the two-jaws law applied inside a single slot, mid-deliberation, and it needs no answer key: consistency is checkable by the solver alone, which means the collider is something the machine can deploy, not only something we can measure. The key grades the collider afterward; it never enters it.
There is a second reading of the pun, and it is the same machine seen from the failure side. Spell it Gödel collider and you get the logic-nerd joke: smash axioms together until the paradox appears. Smash a parse against itself — drop one committed factor at a time and re-propagate — and the contradiction names its own participants, the smallest set of commitments that cannot coexist. Those slots are exactly the ones to re-read. The godol collider tells you which reading survives; the Gödel collider tells you where to collide.
Almost every part of this exists. The bridge between the jaws has a one-call propagator that returns forced facts, or a contradiction. The problem generator has a uniqueness gate that can say whether a diagram forces exactly one answer. And the ledger holds the collider’s ancestor: withhold-and-solve, where the machine withholds the piece of the diagram it trusts least and lets the solver re-derive it from everything else. That trick recovered 26% of what would have been wrong answers, for free, and introduced not one silent wrong answer at any depth, because a re-derived value that contradicts the rest of the graph is refused rather than guessed. The collider is the same idea with the last step generalized: instead of withholding one piece, enumerate its alternatives and let the solver grade them.
One honest correction belongs here. An earlier post on this site said the solver “can hand back a minimal unsatisfiable core.” The core’s search can be made to do that, but no organ in the deployed solver is built and named for it. The Gödel side of the collider needs one, and the cheap version — delete one factor, re-propagate, repeat — is a few lines over the propagator we have. It gets built as part of this, and the earlier sentence gets to become true.
Consistency is not truth. An exactly-determined problem can carry a consistent wrong reading: swap two quantities in a system with no redundancy and you may get a different problem that solves perfectly well. So the collider does not return an answer. It returns a survivor set, and the rule of the fingerpost governs what happens next: one survivor, and the slot is disambiguated; several survivors, and the machine stays silent, because silence earned by counting is worth more than a guess.
That limit is also the first thing to measure, and it is a number nobody has yet read: of the wrong readings our machine produces on wild text, what fraction are self-contradictory at all? If most wrong readings are consistent, the collider is blind to them and disambiguation has to come from somewhere else. We call that number collider visibility, and we pinned it before looking: at least 30% of wrong parses must be contradictory for the collider to be worth building into training, and below 10% we call it dead.
The read is queued behind the run currently on the card, and it costs no training at all: take the champion’s first-pass readings of two thousand wild problems, enumerate the second-best alternative at every torn slot, collide the original and every variant, and count. Flip a reading only when the original is contradictory and exactly one variant survives with a unique answer. Then, and only then, grade the flips against the key. Pinned: visibility of at least 30%; flip precision of at least 0.60; net recovery of at least 5% of the wild failures, with new silent-wrong answers held under 1% — the ancestor managed zero.
If the bars hold, the build is a veto channel: the collider’s verdicts ride into the next cycle of deliberation as conditioning, on the same road the solver’s facts already travel, never as a training signal on the head’s confidence — a monitored signal in the loss teaches concealment, not cure. Trained under this week’s pressure cooker, the machine’s committed channel would learn from collisions. If the visibility bar fails, the Gödel half still earns its keep as an autopsy instrument: the paradox that names which slot was wrong.
Two sages walk into an accelerator. What comes out is a smaller set of things that could possibly be true. That is what reasoning was always supposed to be.