The Pattern Seam · a portal across four layers

Something Has to Stop the Loop

A page matters if pages that matter point to it. A word means what its definition says, and the definition is made of words. Both are definitions that eat themselves, and one of the layers under this page already says what separates them from Gödel's sentence. What none of them says is how many answers the circle has, and that is not a mood, it is a count. Here it is counted twice: on a real dictionary's 1,200 words and 5,607 lookups, where the answer is unique or forty-seven-fold depending on a rule about forty-seven words that nobody writes down; and on self-referential sentences in arithmetic, where a decision procedure sorts every formula up to a size bound into classes and reports exactly how many solve the definition. One answer, none, or nine. The condition that decides it is the same idea in both, and it is not the famous number.

The sentence this page starts after

The layer on PageRank already relates itself to the layer on Gödel, and the note on that edge is good enough that this page has to begin by quoting it rather than pretending to it:

Both start from a sentence that talks about itself — but they end in opposite places. Gödel's self-reference ('this statement is unprovable') breaks the system open: no consistent theory can settle it. PageRank's self-reference ('important if the important point to you') does the reverse: the circle closes into a unique, computable fixed point. Self-reference as a trap, and self-reference as an answer.

the relates: card of The Importance That Points at Itself, published 2026-07-11

Trap or answer is a fork with two branches. The real fork has three, and the third is the one that does the work: a circular definition can have exactly one answer, it can have none, and it can have many. Which of the three you are looking at is a property you can compute, in both settings, and the rest of this page computes it.

The trick is to stop reading the definition as a sentence and start reading it as a walk. "A word is defined by the words in its definition" is a directed graph, and looking a word up is a step along an edge. "A page matters if pages that matter point to it" is the same graph with a different label. Once it is a walk, the question how many answers has a mechanical form, and it is this one:

Is there exactly one place the walk cannot leave?

A set of states you can enter and never leave is a closed class. If there is exactly one, the walk has exactly one long-run answer and you reach it from anywhere. If there are several, there are several answers and which one you land on is decided by where you began, which is another way of saying the definition did not decide it. The count of closed classes is the multiplicity of the eigenvalue 1, so it is a number you can read off the operator rather than a judgement you make about it.

A dictionary, and forty-seven words that point nowhere

The graph is not built for this page. It is the one The Dictionary That Eats Itself ships: the 1,200 most frequent single-word content headwords of Webster's Revised Unabridged Dictionary (1913, public domain), each reduced to its first-sense gloss trimmed to thirty tokens, with an edge from a word to every vocabulary word its gloss uses. This page rebuilds it by slicing that layer's own parse code out of that layer's own verifier and running it over that layer's own lexicon, then checks that every figure the layer publishes comes back: …

Onto that graph goes the circular definition from the other layer. A random reader starts at a word, follows a word out of its definition, follows a word out of that one, and keeps going; with probability 1 − d at each step they give up and jump to a word chosen at random instead. The fraction of time spent on each word is the answer, and d, the damping factor, is the dial everyone knows about.

There is a second dial, and it is not on the instrument panel of any search engine. Of the 1,200 words, 47 have glosses that mention no other word in the vocabulary. The dictionary layer already discloses them and already says what they are: not grounded primitives but artefacts of the 1,200-word slice, words like data, user, computer, whose real defining words fell outside the cut. The walk arrives at one of them and there is nowhere to go. Somebody has to say what happens next, and every published treatment says something different:

1. Spread the mass uniformly. The walker at a dead end jumps to a word chosen at random from the whole vocabulary. This is the standard stochasticity adjustment, and it is not part of the definition of PageRank; it is a repair applied before the definition can run. the usual patch
2. Let the word define itself. A word whose definition mentions nothing else is taken to mean itself. Defensible, and the smallest possible edit to the graph. the smallest edit
3. Let the mass leave. The walk simply ends there and the total stops adding to one. the honest refusal to patch

Take the dial and the rule. The readout below is computed in your browser from the shipped edge list, every time you move anything.

What the bench says, in one line

Under the standard patch there is one closed class at every setting of the dial, all the way to d = 1 with the teleport switched off entirely, and the four starts agree to about 2 × 10⁻¹⁵. Under the self-loop rule there are forty-seven, and at d = 1 the four starts reach four different answers, with 2 of the six pairs at an L¹ distance of 2.000, the maximum, meaning those two share no word at all. The reason is worth seeing plainly: start the walk on path, a word whose gloss mentions nothing else, and under this rule it defines itself, so the walk arrives and stays. The answer is "path, with probability 1.000". Start on data and the answer is data. Same graph, same algorithm, same dial. The count moved because of a rule about forty-seven words, and that rule is not in the equation anyone writes down.

The dial does something else, and it is worth separating. The theorem here is Haveliwala and Kamvar's (2003): the second eigenvalue of the Google matrix satisfies |λ₂| ≤ d, with equality exactly when the underlying walk has two or more closed classes. That is not a bound this page assumes, it is a prediction this page can check, and the check comes out on the nose. With the self-loop rule, where there are forty-seven closed classes, the measured second eigenvalue is 0.850000000 at d = 0.85, equal to d to every digit the measurement carries. With the uniform patch, where there is one, it is 0.512667, six tenths of the bound. So the famous dial is an upper limit on how slowly the loop can settle, and the unwritten rule decides whether it settles on one thing at all.

The knot is the sink, once the artefacts are gone

The dictionary layer's headline object is its Core: the largest strongly connected component of the definition graph, 724 words that only define each other, which 98.6 per cent of lookups out of a Core word never leave. That layer proves the Core exists. It does not ask whether the Core is where the loop's answer lives, and the answer is almost, and then exactly.

Almost, in the raw graph: the Core is not a closed class, because a handful of its edges run out to the artefact words, and it carries 91.31 per cent of the answer at d = 0.85 rather than all of it. Exactly, once the artefacts are removed the only way that does not require choosing a patch: delete every word that points nowhere, then delete every word that now points nowhere, and keep going until nothing is left to delete. That removes 48 words and leaves 1,152. On what is left there are 1 closed classes, the single one is the 724-word Core itself, and with the teleport switched off it holds 100.0000 per cent of the answer.

That is the whole shape of the thing in one sentence. A dictionary's definitions form a knot you cannot get out of by looking words up, and if you ask which words matter by a rule that only ever follows definitions, every last unit of the answer ends up inside the knot. The teleport is not a technical convenience for handling awkward graphs. It is the only thing in the model that comes from outside the dictionary, and without it the dictionary's answer about itself is entirely about itself.

Is "important" the same as "must be grounded from outside"?

The temptation at this point is strong and should be resisted in public. The dictionary layer carries a second object, the minimum grounding set: the fewest words you would have to learn from outside language for the rest to become reachable, which it formalises as a minimum feedback vertex set and which is NP-hard, so that layer ships a greedy upper bound of 465 words. This page has just argued that the teleport is the place where something from outside enters. Two "you must come in from outside" sets, on one graph. It would be a lovely sentence to write that they are the same object.

They are not, and the gap is measurable. Take the top 465 words by importance under the standard patch at d = 0.85 and intersect them with that layer's greedy grounding set of the same size:

Well above chance, nowhere near identical. Being important and being load-bearing are related questions about the same knot and they are not one question, which is worth knowing because the cheap move here would have been to run one and report it as the other.

One case can be settled exactly rather than heuristically, and it goes the other way. On the six-word subgraph the dictionary layer brute-forces, cause, effect, result, act, power, make, the minimum grounding sets are exactly three, all of size two: {cause, act}, {cause, power}, {act, make}. Order those six by importance on the full graph and the top two are act, make, which is one of the three. So on the one instance where the exact answer is known, importance picks a genuine minimum grounding set, and on the full graph the same heuristic is right about 66.2 per cent of the time. Both facts are the finding; neither on its own is.

The same question, asked in arithmetic

Now the other half. In arithmetic a definition that mentions itself is a sentence, and the classical fact is that you can always build one: for any property you like there is a sentence asserting that property of itself. The Sentence That Says It Can't Be Proved makes that construction operable, and it is careful to claim only that such a sentence exists. It never says how many there are.

There is a theorem about that, and it has the same shape as the closed-class count. Work in GL, the modal logic whose □ reads "is provable in Peano arithmetic" and whose one non-trivial axiom is Löb's, and which Solovay (1976) proved captures exactly the principles of provability that arithmetic can prove about itself. A definition is a formula A(p); an answer is a formula F containing no p with GL ⊢ F ↔ A(F). De Jongh and Sambin, independently, in 1975:

Suppose that all occurrences of the propositional variable p in a given formula A(p) are under the scope of the provability operator… Then there is a formula B in which p does not appear, such that all propositional variables that occur in B already appear in A(p), and such that GL ⊢ B ↔ A(B). Moreover, the fixed point is unique, or more accurately, if there is another formula C such that GL ⊢ C ↔ A(C), then we must have GL ⊢ B ↔ C.

Stanford Encyclopedia of Philosophy, "Provability Logic", stating the de Jongh and Sambin fixed point theorem

The condition is called being modalized in p, and in plainer words it is a guard: every mention of the sentence inside its own definition has to pass through one more layer of provability. It is the exact analogue of the closed-class question, and the count it predicts is the same count. So do not take the theorem's word for it. Below, the page enumerates every formula up to size five built from ⊥, ¬, □, ∧, ∨ and →, sorts them into GL-equivalence classes with a tableau decision procedure, and then asks of each class whether it solves your definition. The number of classes that survive is the number of answers.

Every guarded definition on the preset row returns exactly one answer, at every bound large enough to hold it. Every unguarded one returns something else. ¬p, the liar, returns none, and no bound on the search changes that. p, the truth-teller, returns every class in the space, because a definition that says nothing constrains nothing, so its count is just the size of the space and rises when you raise the bound. □p → p, the antecedent of Löb's axiom, returns more than one and not everything: at the default bound it is ⊤ and ¬□⊥, the sentence saying the theory is consistent. A partly guarded loop narrows the answers without picking one.

The last preset is there because the instrument can fail in the way the rest of this page is about. □□□□□¬p is guarded, so exactly one answer exists, and at size five the search returns nothing at all: the answer is □□□□□⊥, which is one symbol too big for the space. A bounded search that comes back empty is not the same statement as "there is no answer", and the solver says which of the two it is by looking at the guard first. Raise the bound to six and the answer appears.

Two details worth the space. The class count is not a design choice: 142 distinct formulas of size 5 or less collapse, under this logic, to exactly 9 things, the tower …. And the answer the solver returns for ¬□p, the Gödel sentence, is ¬□⊥: the statement that the theory does not prove a contradiction. The encyclopedia states that same fixed point in the same words, which makes it a check on this page's prover from outside this page.

Stopping and having one answer are different conditions

Both halves of this page tempt you into the same conflation, so both halves get to refute it.

Ask the definition x = 1 − x over and over from x = 0 and you get 0, 1, 0, 1, forever. The iteration never stops. The equation has exactly one solution, x = ½, sitting there the whole time. Reachability by repeated asking and uniqueness of the answer are two different conditions on the same object, and the second does not imply the first.

The logical half shows the same crack in reverse, and shows it on a case that matters. Ask ¬□¬p repeatedly from ⊤ and the iterates run ¬□⊥, ¬□□⊥, ¬□□□⊥, … and never repeat. The definition is guarded, so by de Jongh and Sambin it has exactly one answer, and the solver above finds it by search: it is ⊥. The iteration simply never arrives.

And there is a third failure the page walked into, which is left in because it is the most useful thing here. The finite frames the tableau agrees with are the standard semantics for GL (Segerberg 1971, completeness for finite transitive irreflexive frames), so enumerating them refutes soundly. But enumerating them only up to a bound validates only up to that bound, and asking a bounded frame checker when the iteration has stopped moving gives a confident answer that changes when you raise the bound:

The two instruments are also crossed directly, on 4,000 pseudorandom formulas from a seeded generator: the tableau and the four-world enumeration agree on 4,000 of 4000, with 0 cases where the tableau called a formula valid that a small frame refutes, which would be a bug rather than a blind spot. And the tableau is held to 8 textbook facts that it could fail, 4 of which it must call false:

Where there is no guard, and no theorem to appeal to

The last member is the case that has none of this machinery available. The Focal Point is Schelling's coordination problem: two strangers who cannot communicate must choose the same thing, so each should choose what they expect the other to expect them to choose. That is a definition that eats itself and it is completely unguarded. Every profile where both choose alike is an equilibrium, so the answers are as numerous as the options and nothing inside the mathematics narrows them. What narrows them, in Schelling's account, is salience, which is not in the game.

That layer put the game to free models, alone and unable to signal, and its measured ladder is the shape of a loop with many answers and no selector:

On the coin the machines coordinate at 71 per cent, close to the human rate Schelling reports. As the choice space opens the agreement falls away, and on a free single word it is zero: every model that answered wrote a different one. Read against the rest of this page, that is not a story about machines being bad at coordination. It is what an unguarded loop looks like when you actually run it: the definition is satisfied by any agreement at all, so it selects nothing, and whatever does the selecting for humans is not in the specification and these minds mostly do not have it.

What this page claims, and what it does not

Claimed. That "how many answers does this circular definition have" is a well-posed question with a computable answer in both settings on this page; that in the numeric setting the answer is the number of closed classes, which on this dictionary is one or forty-seven depending on a convention for forty-seven words that no published statement of PageRank contains; that in the logical setting the answer is one whenever the self-reference is guarded and something else when it is not, verified by exhaustive search over an enumerated space under a decision procedure rather than taken on the theorem's authority; that the Core of this dictionary is the unique closed class of the pruned graph and holds all of the undamped answer; and that importance and grounding, two ways of asking what a knot depends on, overlap at 66 per cent against a 39 per cent baseline and are therefore not the same set.

Not claimed. That these are one theorem. They are two theorems that rhyme. Banach's contraction principle and the de Jongh and Sambin fixed point theorem are about different objects, and the guard in one is a numeric modulus while the guard in the other is a syntactic position. There is a real body of work that does build the bridge, in which a "later" modality gives guarded recursive definitions unique solutions by Banach's theorem on complete ultrametric spaces, and in which Löb's axiom is the guarded fixed point principle: Nakano's "A Modality for Recursion" (2000) and Birkedal, Møgelberg, Schwinghammer and Støvring's "First Steps in Synthetic Guarded Domain Theory" (2011) are the entry points. This page does not make that bridge, does not depend on it, and names it so a reader who wants the formal version knows it exists and knows this is not it.

Also not claimed. Anything about meaning. The dictionary layer hands its Core to Harnad's symbol grounding problem and is careful that the inference is contested; this page stays on the arithmetic of the graph and takes no position on whether meaning can ground in language. And nothing here says the machines in the last section lack a theory of mind; the ladder measures agreement in six games and that is all it measures.

What was already here

A portal earns its place by the claim its members do not make, so the ones they do make have to be on the page.

• The Importance That Points at Itself already publishes the join between PageRank and Gödel, quoted at the top, and its own verifier already notes operationally that the error decay ratio equals |λ₂| and that the theorem bounds it by d. What is new here is running that engine on a corpus rather than a toy graph, and turning the bound into a count.
• The Dictionary That Eats Itself publishes the graph, the Core, the greedy grounding set, and the disclosure that the forty-seven sinks are artefacts. It does not run a random walk on any of it. Its own card relates it to No Fix From Inside, a portal that already holds the grounding argument; this page does not re-open it.
• The Fixed Point is the closest neighbour and the one to read next. It shows that self-reference always manufactures a fixed point and sorts the results by what becomes of them: forbidden, undetermined, unprovable, self-fulfilling, merely true. Its discriminator is which operator you chose; this page's is how many solutions the choice admits, which is a different question with a numeric answer. Its truth-teller card and this page's nine classes are the same phenomenon met from two directions.
• The Limits of Knowing already walks Gödel under another name and already asks one sorting question of several walls. Same move, different question: it asks whether a diagonal is present, this asks how many fixed points the diagonal has.
• The Focal Point's ladder is reproduced here from its own data file and is not re-analysed. A convergence curve for it is composted as a pending improvement to that layer and is deliberately not built here.