The paradox you cannot silence — and the soundness you cannot speak
A theory would like to endorse itself — to assert, in one stroke, everything I prove is true. That single sentence needs a word for its own truth, and the Liar shows the word cannot be had. This study traces the wish to its refusal: the diagonal in its barest semantic form, Tarski's proof that truth is undefinable, the conservation of deficit that makes every repair of the Liar breed a stronger one, and the exact theorem by which a truth predicate — and only one that enters the rules of proof — finally supplies the soundness a system lacks. The climb of the whole work begins here, at the endorsement no single level can give.
the wish · undefinability · conservation of deficit · revenge · truth's office · the first rung
Across the later movements a theory keeps reaching for the rung above it, and the reaching always has the same motive: it wants to say of itself that it is sound — that what it proves is true. The whole ascent is that wish, repeated at ever larger scales. This first movement asks where the wish comes from and why it can never be satisfied in place, and the answer is older and barer than any of the arithmetic that follows. It lives in a single sentence that talks about its own truth.
This is the opening movement of The Ascent, drawn down to its root. We begin with the endorsement a theory would like to give itself and the truth predicate it would need. We prove, by the same diagonal that drives every limitative result, that the predicate cannot exist — Tarski's undefinability of truth, the Liar made exact. We state the invariant the rest of the corpus tracks: that closure, consistency, and effectiveness are three properties no theory of its own semantics can hold at once, and that this joint impossibility simply is Tarski's theorem and Gödel's together. We watch each classical repair of the Liar quiet it in the vocabulary it had and be undone by a stronger Liar in the vocabulary it was forced to add. And we reach the hinge that opens onto the climb: a compositional truth predicate, and only one allowed to enter the rules of proof, proves exactly the consistency and reflection its base theory could not — so that the theory of truth and the theory of reflection turn out to be one ladder, climbed in the movements to come. The endorsement begins as a wish and ends as a measured cost.
A consistent theory believes its own theorems. It would like to say so — to assert, once and for all, that everything it proves is true. The trouble is hidden in the word true: to endorse itself in a single sentence, a theory needs a predicate for its own truth, and that predicate has a sentence waiting for it.
Imagine a theory T — arithmetic, say — trying to vouch for itself. Not for this theorem or that one, but for all of them at once: whatever I prove is true. To write that down as a sentence of its own language, T needs a one-place predicate True(x) that applies to (codes of) sentences, so that it can say ∀φ\u2009(Prov_T(⌜φ⌝) → True(φ)). The endorsement is exactly this: a single closed sentence asserting the soundness of the whole theory.
Whatever else a truth predicate does, it must disquote: for every sentence φ,
And here the oldest paradox in logic steps forward. Consider the sentence that says of itself that it is not true — this sentence is false. It is the diagonal in its barest semantic form: the truth predicate supplies the self-naming, negation supplies the twist, and the sentence is the fixed point the construction forces and forbids in one breath. The endorsement wants a truth predicate; the Liar shows what any such predicate must contain. The next section makes the collision exact.
The collision is a theorem, and its proof is three lines. The same diagonal lemma that powers every limitative result, aimed at the negation of a truth predicate, produces the Liar as a fixed point — and the disquotational schema turns that fixed point into the contradiction L ↔ ¬L. No consistent theory can define its own truth.
The engine first — the construction underneath Gödel, Tarski, and the rest.
Now suppose, for contradiction, that T did possess a truth predicate True(x) satisfying every instance of the schema. Apply the diagonal lemma to ¬True(x). It hands back a sentence L — the Liar — with T ⊢ L ↔ ¬True(⌜L⌝). But the schema, at φ = L, gives T ⊢ True(⌜L⌝) ↔ L. Chain the two equivalences and True(⌜L⌝) cancels:
The endorsement asked for a predicate that disquotes every sentence. That is precisely the predicate Tarski's theorem forbids. The single sentence by which a theory would vouch for all its truths cannot even be written in the theory's own language — not because it would be false, but because the word it needs does not exist there. The soundness a theory most wants to state is the one statement it is structurally barred from making.
This is the precise shape of “truth is always one level up.” Each level's truth is in hand; what no level holds is the truth of the whole language including itself. The refusal sits not at any rung but at the limit — which is the structure the next sections exploit, and the one the reflection ladder will later climb.
Tarski's theorem is not an isolated curiosity. Together with Gödel's it states a single conservation law: no theory of its own semantics can be at once closed, consistent, and effective. The three are jointly unsatisfiable. Every repair of the Liar buys two of them by surrendering the third — and the deficit itself is what survives the repair.
Name the three demands. A theory is closed if it defines and decides its own truth — contains a predicate that disquotes every sentence and settles each one. It is consistent if it proves no contradiction. It is effective if its axioms are recursively enumerable — a machine can list them. Each pair is attainable; the three together are not, and the obstruction is exactly the two theorems we have been circling.
PA, ZFC. A recursively axiomatized consistent theory cannot define or decide its own truth (Tarski, Gödel). It proves true things; it cannot say, inside itself, that they are true.
the working foundationsTrue arithmetic Th(ℕ). The set of arithmetic truths is consistent and semantically complete — but not recursively axiomatized, not even arithmetical. No machine generates it.
truth itself, ineffableThe naïve truth theory: PA plus the full disquotational schema. It defines and decides its own truth and is recursively axiomatized — and it is inconsistent, proving L ↔ ¬L.
the paradox, embracedthe semantic root of incompleteness And the two theorems are not merely siblings; one yields the other. Provability in a recursively axiomatized theory is Σ⁰₁ — definable, listable by a machine. Truth, by Tarski, is not even arithmetical. Suppose a sound, recursively axiomatized theory were also complete, deciding every sentence. Then for each φ, soundness makes T ⊢ φ hold exactly when φ is true, so Th(ℕ) would coincide with the provable sentences — and would therefore be Σ⁰₁, definable, contradicting Tarski. No sound, effective theory is complete. That is Gödel's first theorem, falling straight out of the undefinability of truth. Provability is a definable approximation to a target that is provably beyond definition, and the gap between them is forced — the Gödel sentence its first witness.
The temptation is to think one of the classical treatments solves the Liar. Watch instead what each actually buys, and — more tellingly — what it can no longer say. The pattern is exact: a repair silences the Liar in the vocabulary it had, and the strengthened Liar is written in the vocabulary the repair was forced to add. The paradox is relocated, never removed.
Each repair is a choice of which two of the three demands to keep. Select one and watch the third give way — and watch where the next Liar appears.
the most instructive escape Kripke's is worth seeing in full, because it comes closest. Start with the truth predicate undefined everywhere, then apply a rule: declare True(⌜φ⌝) true once φ has been made true, false once φ has been made false, and leave it undefined otherwise. Iterate the rule, transfinitely, and it climbs to a least fixed point at which True(⌜φ⌝) and φ always share their status. There the grounded sentences — those that bottom out in non-semantic fact — receive classical values, and the ungrounded ones, the Liar foremost, fall into a truth-value gap: neither true nor false. The Liar is defused, not as a contradiction but as a sentence with no value. A single self-applicable truth predicate survives — at the price of bivalence.
But now the strengthened Liar: this sentence is not true. In the gapped language, “not true” can only mean “false-or-gapped,” and the predicate that expresses that — the predicate “is gappy” — is not in the object language; it lives in the classical metalanguage that surveys the fixed point. The moment one can say “L is not true,” the paradox is back, one level up, in exactly the vocabulary the repair had to step outside itself to use. Tarski's hierarchy reappears at the level of the gap. The determinacy operator that tries to internalize “is gappy” only breeds its own Liar and must be iterated through the ordinals to chase it — which is the reflection ladder of the next movements, breaking surface inside the theory of truth itself.
Tarski forbids a theory its own truth predicate. He does not forbid a fresh one, added carefully and typed to the language below. The remarkable fact — the hinge onto the entire ascent — is exactly how much such a predicate buys, and that the answer turns on one thing: whether truth is allowed to enter the rules of proof.
Step outside the object language, as Tarski's hierarchy requires, and add a new predicate T(x) meant to apply to the sentences of the base theory PA — and to those only, never to sentences containing T itself. This typing sidesteps the Liar: there is no sentence of the right kind to diagonalize. The question is what the addition is worth, and the answer is a clean dichotomy that the strength ladder below lays out.
Give T the full compositional theory — Tarski's clauses, T commuting with negation, conjunction, the quantifiers — but keep induction restricted to formulas not mentioning T. The resulting theory CT⁻ is conservative over PA: it proves not one new arithmetical theorem (Kotlarski–Krajewski–Lachlan, 1981). A complete inductive account of truth, and arithmetically inert. Now lift the one restriction — allow induction over formulas that mention T — and the theory CT proves the global reflection principle every theorem of PA is true, and with it Con(PA). The soundness the base theory could not state, it now possesses.
The dichotomy is the lesson. A truth predicate that merely disquotes is a convenience; a truth predicate that participates in proof is strength — and it is precisely the strength the base theory was missing, its own reflected soundness. This is why the next movements can be conducted in the language of reflection alone, with no mention of truth, and lose nothing: truth's whole office, on the arithmetical side, is to supply the soundness the system lacked (Tarski; Feferman's reflective closure). The endorsement is available after all — not for free, and not from inside, but as a measured ascent whose cost the rest of the corpus reads off to the ordinal.
Everything that follows is the recovery, level by level and never all at once, of the endorsement this movement showed to be impossible in place. The semantic refusal and the syntactic one are a single theorem; the ladder the corpus climbs is the staircase that refusal forces.
Set the two faces of the obstruction side by side. Tarski: a consistent theory cannot define its own truth. Gödel: a consistent effective theory cannot prove its own consistency. They are the same diagonal, run on the same fixed-point lemma, refusing the same thing — a system's complete grip on its own soundness. One refusal is semantic, about truth; the other syntactic, about proof; and §III showed that the first delivers the second. A theory cannot hold its own endorsement, in either currency.
A consistent theory cannot contain a predicate that disquotes all its sentences. Truth for level n is definable only at level n+1.
A consistent effective theory cannot prove its own consistency. Soundness for theory T is provable only in a strictly stronger theory.
We began with a wish — that a theory might say, in one breath, that all it proves is true — and found the breath unsayable: the word true carries a sentence that turns the saying into a contradiction. That is Tarski's theorem, and it is Gödel's, and it is the conservation of a single deficit that no repair removes but only relocates, always to the vocabulary the repair was forced to add. Yet the endorsement is not simply lost. Stepped outside the theory and admitted into its proofs, a truth predicate supplies exactly the soundness the theory lacked — and the price of that supply, measured to the ordinal, is the ladder the rest of this work climbs. The first rung is the oldest sentence in logic, the one that speaks of its own truth. From it the whole ascent unfolds: not a tower built on a foundation, but a staircase forced by the one endorsement no single step can give itself.