Incompleteness Without Infinity

1. Introduction

The question.

The incompleteness theorems are the most frequently cited results of modern logic, and among the most frequently deployed outside it. Within the philosophy of mathematics their standard employment is evidential. Gödel’s first theorem is read as the obituary of Hilbert’s programme [1,2], as the demonstration that mathematical truth outruns every formal grasp [3], and, in its most ambitious uses, as a proof that minds are not machines [4,5]. In each employment the theorem is taken to bear against positions that would confine mathematics to the finite or the formal: against finitism, against formalism, against mechanism. Gödel himself drew a platonist moral. The pattern of these readings is constant. The theorem exhibits a truth beyond the reach of a given apparatus, and the excess is credited to the world: mathematical reality is larger than any effective description of it.

This paper asks what becomes of the phenomenon if the ontology of mathematics is finite. The question is not rhetorical. A finite ontology, under which every mathematical structure that exists is finite and the appearance of inexhaustibility is a horizon effect of bounded agency, is an articulated position with a developing literature [6,7,8]. Whatever its merits, the position is entitled to a precise answer to the natural challenge: does incompleteness refute it, embarrass it, or bear on it at all?

Thesis.

Three claims, in ascending strength.

First, vacuity: over a fixed finite structure, Gödel’s hypotheses fail. The structure interprets no arithmetic whose iteration is endlessly fresh, the successor injective and never recurring, its first-order theory is complete and decidable, and there is no true-but-undecidable sentence of the structure (Section 3). This much is elementary, and the paper claims no originality for it. Its role is to set the stage for what finitude forces next.

Second, decomposition: the phenomenon survives, but not wholesale, and how it survives is the paper’s finding. A proper part of a finite structure cannot faithfully represent the whole; this is a cardinality theorem (Section 4). An agent with a bounded certification budget reaches only a vanishing fraction of the structure’s truths, and every truth beyond its length horizon outright (Section 5). These counting facts are the component of incompleteness that migrates as theorems. A second component concentrates in proof complexity and is proved for specific systems, open in general: whether short truths defeat every locally checkable system, exactly parallel for the first theorem and the second. And a third component does not transfer: a no-compression theorem shows that the diagonal, the engine of the classical proofs, cannot be written over a fixed finite structure, wholly for per-symbol coding and beyond a structure-dependent threshold for any coding, the remainder an open definability problem, because it consumes exponential compression of quotation, a resource infinity supplies and finitude withholds; correspondingly, bounded-fragment truth becomes definable at polynomial overhead, and Tarski undefinability reverses (Section 5). Incompleteness moves from the truth axis to the reach axis, and the move is a decomposition: counting facts, complexity problems, and one instructive impossibility.

Third, neutrality: because the decomposition preserves the certificates, incompleteness cannot adjudicate the ontology debate. Every epistemic contact a bounded agent can have with the classical limitations, a failed search, a schematic derivation, a lower bound for its own system, a conditional metatheorem, and every agent is bounded, is available verbatim inside a sufficiently large finite world; the detached totality-level facts are available to neither agent. The difference between the ontologies is not registered on any certificate either agent can possess. The theorems measure the finitude of the knower, not the size of the world (Section 7). The classical reading, undecidability of truth as a property of mathematical reality, is diagnosed as the shadow cast by one particular idealization: the demand that an infinite domain be governed by an effective description. Withdraw the infinity and the same gap reattaches to the agent, where it had operational content all along.

What is not claimed.

The paper does not argue that the ontology of mathematics is finite; that argument belongs to a different literature and is cited where used, never presupposed. It revises no logic: classical logic and standard metatheory are in force throughout. It makes no feasibility or vagueness claims: no sorites is run, no human limitation is invoked, and the agent bounds below are parameters, not psychology. And it disputes no classical derivation. Every result of Gödel, Tarski, and their successors is retained as proved, the incompleteness and undefinability proofs being themselves finitary syntax, and the paper uses several of them: derivations are finite acts on finite strings, and nothing in what follows questions an act. What the paper declines to add is an interpretive layer the derivations do not carry, the assertoric, totality-level reading of statements that quantify over an infinite domain; Section 7 prices that reading as an idealization to be read schematically, and the neutrality argument never needs it. Where the paper evaluates infinite structures directly (Theorem 5), it says so and argues by that ontology’s own semantics, as an elimination argument must. What is disputed outright is an evidential employment: the inference from those theorems to conclusions about the size of mathematical reality.

What is proved and what is diagnosed.

Section 2 through Section 6 are mathematics: elementary, but with every hypothesis displayed, since the displayed hypotheses carry the philosophy. The results are stated for an arbitrary fixed finite structure; a single running example (modular arithmetic on a prime) supplies exact magnitudes. Section 7 is philosophy argued from the mathematics: the neutrality claim, the diagnosis of the classical reading, and the consequences for the standard employments of the theorems. Section 8 answers objections. A reader who rejects the finite ontology loses nothing of Section 2 through Section 6 and retains the neutrality claim in conditional form: whichever ontology is true, incompleteness does not say which.

2. Preliminaries: Structures, Agents, and Grades of Decision

Structures.

Throughout, $\mathcal{M}$ is a structure in a finite signature $\sigma$ with $\left|\mathcal{M}\right|=m$ finite, $m\ge 2$. The first-order language of $\sigma$ is L; $\mathrm{Th}\left(\mathcal{M}\right)$ is the set of L-sentences true in $\mathcal{M}$. Where a result needs a specific feature of $\sigma$ (a binary operation, two constants) the hypothesis is stated in the result. The running example is ${\mathcal{M}}_p=(\{0,1,\dots ,p-1\};0,1,+,\times )$, arithmetic modulo a prime p: a structure rich enough to carry non-trivial algebra, poor enough (by finitude alone) to escape Gödel’s hypotheses. Nothing in the paper depends on the choice of example.

Iteration and induction.

One terminological rule prevents a family of misreadings. Iteration names the successor act: given a number, form its successor. The act is not what an infinite arithmetic asserts. A finite cyclic structure licenses the act without bound: successor is total on the cycle, every application is defined, no attempted step fails; the act wraps. What an infinite arithmetic asserts is the act’s endless freshness: successor injective and never returning to zero, every application reaching a new entity. Throughout the paper, “unbounded iteration” abbreviates this freshness claim, iteration asserted to be endlessly non-recurrent, and never the act itself. Induction names the proof principle: what holds of 0 and is preserved by successor holds of every number. The two are independent, in both directions. Robinson arithmetic $\mathsf{Q}$ has unbounded iteration and no induction axiom at all, and every model of $\mathsf{Q}$ is infinite (Proposition 1); a finite linear order satisfies the least-number principle, and hence induction, as a theorem, while licensing no unbounded iteration. The infinitary commitment of arithmetic tracks iteration exactly and induction not at all. Tracks it, note, at the level of model size: unbounded iteration is what forces every model to be infinite. What drives incompleteness is a further ingredient on top of iteration, coding power, and Section 6 separates the two with a complete, decidable theory whose iteration is unbounded. A finite ontology therefore keeps induction, and keeps the act; what it declines is the freshness claim. Readers for whom finitism connotes suspicion of induction should read the present position as its inverse.

Agents and certificates.

The epistemic subject of the paper is a bounded agent, modelled with deliberate austerity.

Definition 1 

(Agent, certificate, retention). Fix a finite alphabet A with $\left|A\right|=a\ge 2$. Anagentof capacity $(K,H)$ possesses at most K symbols of storage and performs at most H elementary steps in any run; an elementary step reads or writes one symbol or evaluates one atomic fact of $\mathcal{M}$ (one entry of an operation or relation table). Fix also aderivation system$\mathsf{D}$: a finite set of inference rules over L-sequents whose individual steps arelocally checkable, meaning verifiable in time polynomial in the step’s written length, the polynomial denoted $p_{\mathsf{D}}$ and called the checking polynomial of $\mathsf{D}$, with each lookup in the tables of $\mathcal{M}$ counted as one step, and which issound: every $\mathsf{D}$-derivable sentence is true in $\mathcal{M}$. Formula and record length is throughout measured as string length over the fixed finite alphabet, variable indices written out in the object syntax; a formula with k distinct variables therefore spends order $klogk$ symbols on names alone once k exceeds the alphabet. The local-checkability clause keeps two resource measures from being conflated: the length of a derivation (what certificates carry) and the time to verify it (what budgets charge). Without the clause, a rule may hide unbounded verification work inside a short step, and Proposition 5 shows the hiding is fatal to any uniform claim about short underivable truths. Acertificatefor a sentence φ is a record, retained in the agent’s storage at the end of a run, that encodes a $\mathsf{D}$-derivation whose designated end sentence is φ; a record certifies exactly the one sentence its derivation ends with. An agentcertifiesφ at budget H if some run of at most H steps writes the record, step-checks it against the rules and the tables (the cost priced by the local-checkability clause), and retains it: certification includes verification, and a transcribed but unchecked record certifies nothing. Retention is part of the definition: an output written and discarded is not possessed, and a fact once verified but no longer held is not certified. Since a certificate is retained in storage and written within the run, its length is at most $Bmin(K,H)$, the agent’s effective certificate bound.

The model is neutral between ontologies: nothing in Definition 1 requires $\mathcal{M}$ to be finite, and the same agent can be placed inside an infinite structure. This neutrality is used in Section 7.

Grades of decision.

The argument requires one distinction held steadily: what it is for a sentence to be decided. Three grades.

Definition 2 

(Grades of decision). Let φ be a sentence with a determinate truth value in $\mathcal{M}$.

• φ isdecided at grade (i)if it has a truth value and nothing more: no presentation of a decision, of any size, is exhibited. Bivalence by fiat.
• φ isdecided at grade (ii)if a decision of φ is specified as a finite combinatorial object over the structure’s tables: the exhaustive evaluation of φ over $\mathcal{M}$, a tree of quantifier instantiations of size exactly computable and on the order of $m^q$ for φ of quantifier depth q. What grade (ii) asserts is the specification and the computed size, not the realization of the object anywhere; the specification and the size computation are themselves short records (Objection O8, Section 8).
• φ isdecided at grade (iii)for an agent if that agent certifies φ within its budget (Definition 1).

Thereach gapof $(\mathcal{M},\mathsf{D},K,H)$ is the set of sentences decided at grade (ii) but not at grade (iii): truths of the totality beyond the agent’s reach.

Grade (ii) is the finite ontology’s rendering of “the totality decides every sentence,” and it is stronger than grade (i) in exactly the way that matters: the decision is a finite object with a size, not a truth value postulated in the abstract. The classical theory $\mathrm{Th}\left(\mathbb{N}\right)$ of true arithmetic is the paradigm of grade (i): complete and consistent, but effectively presentable by nothing and no one, a point Section 6 makes precise. The entire epistemic action of this paper, on either ontology, happens between grades (ii) and (iii).

3. Gödel’s Hypotheses Over a Finite Structure

The first incompleteness theorem, in its sharp modern form, states: every consistent, effectively axiomatized theory that interprets Robinson arithmetic $\mathsf{Q}$ is incomplete [9,10,11]. Three hypotheses: consistency, effectivity, interpretation of $\mathsf{Q}$. This section shows the third fails over any fixed finite structure, while the properties the theorem forbids from coexisting are all present.

Proposition 1 

(No interpretation of $\mathsf{Q}$). No finite structure interprets $\mathsf{Q}$.

Proof. 

$\mathsf{Q}$ proves that the successor function is injective and that 0 is not in its range; thus in any model the elements $0,S0,SS0,\dots$ are pairwise distinct, and every model of $\mathsf{Q}$ is infinite. An interpretation of $\mathsf{Q}$ in a structure $\mathcal{M}$ defines a model of $\mathsf{Q}$ on a quotient of a definable subset of some ${\mathcal{M}}^k$; if $\mathcal{M}$ is finite so is any such quotient, and a finite set carries no function that is injective yet omits an element from its range: injectivity forces surjectivity on a finite set by pigeonhole. □

Proposition 2 

(Completeness and decidability). For finite $\mathcal{M}$: the theory $\mathrm{Th}\left(\mathcal{M}\right)$ is complete, consistent, and decidable, and $\mathcal{M}$ is characterized up to isomorphism by a single first-order sentence ${\sigma }_{\mathcal{M}}$.

Proof. 

Completeness and consistency are immediate: truth in a fixed structure is bivalent. Decidability: a sentence $\phi$ of quantifier depth q is evaluated by exhaustive search, each quantifier ranging over the m elements, in at most $c{m}^q$ atomic evaluations for a constant c depending on the length of $\phi$. The characterizing sentence ${\sigma }_{\mathcal{M}}$ asserts the existence of exactly m pairwise distinct elements, records the complete tables of every operation and relation on them, and asserts that every element is one of them; any model of ${\sigma }_{\mathcal{M}}$ is isomorphic to $\mathcal{M}$. □

Remark 1 

(Trakhtenbrot fenced). Trakhtenbrot’s theorem, that validity over the class of all finite structures is not recursively enumerable [12,13], concerns the class, quantified over all finite models at once. A finite ontology in the sense examined here commits to a structure, not to the class; for a fixed $\mathcal{M}$, Proposition 2 applies and no undecidability enters. The distinction recurs: it is precisely the difference between a totality and a quantification over possible totalities.

Corollary 1 

(Vacuity). Over a fixed finite structure $\mathcal{M}$, neither incompleteness theorem has an instance. $\mathrm{Th}\left(\mathcal{M}\right)$ is consistent, effectively axiomatizable (it is decidable, hence axiomatizable by itself), and complete; no contradiction with Gödel arises, because the third hypothesis, interpretation of $\mathsf{Q}$, fails by Proposition 1. There is no sentence of L true in $\mathcal{M}$ but undecided by $\mathrm{Th}\left(\mathcal{M}\right)$, and no analogue of the unprovability of consistency arises at the level of the totality.

Remark 2 

(Tarski scoped). Tarski’s undefinability theorem [14,15] concerns a structure rich enough to code its own syntax. A finite $\mathcal{M}$ is not: L has infinitely many sentences and $\mathcal{M}$ has m elements, so no injective Gödel numbering of all sentences into $\mathcal{M}$ exists, for cardinality reasons alone. The truth predicate of $\mathcal{M}$ is definable externally, indeed given by a finite table together with the evaluation procedure of Proposition 2; what fails is only its internalization, and Section 4 locates that failure where it belongs, at the part–whole boundary. Undefinability of truth, like undecidability, will reappear as a statement about bounded parts.

The stage is now set. A finite totality is complete, decidable, and immune to the diagonal in its classical form. If that were the end of the matter, finitude would have purchased a suspicious epistemic paradise: everything true would be, in principle, transparently so. The next two sections show the paradise does not exist. The limitative role of the diagonal relocates to every bounded agent; the diagonal mechanism itself, Section 5.2 will show, does not survive the move.

4. The Part Cannot Hold the Whole

The first migrated limitation is representational and static: it depends on no budget, no derivation system, and no notion of time. It is a statement of cardinality.

Theorem 1 

(Relational diagonal). Let $\mathcal{M}$ be finite with $\left|\mathcal{M}\right|=m$ and let $\mathcal{O}$ be an agent of storage capacity K over alphabet A, $\left|A\right|=a$, with $a^K<m$. Then:

• there is no injective encoding of the elements of $\mathcal{M}$ into the agent’s storage states; every internal representation of $\mathcal{M}$ identifies distinct elements;
• a fortiori, there is no faithful self-inclusive internal model: no representation of $\mathcal{M}$ within $\mathcal{O}$ that also represents, faithfully, the encoding map by which it represents.

Proof. (1) The agent’s storage takes at most $a^K$ distinguishable states; a representation distinguishing all m elements would inject $\mathcal{M}$ into the state set, impossible when $a^K<m$ by pigeonhole. (2) A faithful self-inclusive model must represent, beyond the m elements, the graph of its own encoding map, one pair per element of $\mathcal{M}$: the objects to be distinguished number at least m already, and the case reduces to (1). The point of stating (2) separately is not the count but the shape: the obstruction is met by the representation map itself, the agent’s own act of representing turning up among the things it must represent. The shape recalls the diagonal, but only the shape: whether the diagonal itself survives finitude is settled, negatively and instructively, in Section 5.2. Note also the theorem’s scope: it bounds injective, element-distinguishing representation; generative or compressed descriptions are not excluded, and holding one is holding a recipe, not the model (Remark 5). □

Remark 3 

(The limitation, relocated). Classically the diagonal argument, run inside a theory that codes its own syntax, produces a true sentence the theory cannot prove. Over a finite ontology the same self-application, run against a bounded part, produces a whole the part cannot encode. The limitation has changed address, from provability of truths to representability of the totality, and in its representational form the relocation is complete: a theorem about every possible internal representative, obtained by counting, free of any assumption about the agent’s procedures, intelligence, or lifetime. The present remark claims exactly that representational face and no more; what becomes of the diagonal mechanism itself is Section 5.2’s question, answered there in the negative. Tarski undefinability relocates in its representational face the same way (Remark 2): truth of $\mathcal{M}$ is externally a table, internally unrepresentable, because the internal definer is a part and the syntax outruns it. Theorem 1 delivers this as a representational bound under its counting hypothesis $a^K<m$, and the bound dissolves for agents large relative to the world; what becomes of the undefinability itself is Section 5.3’s reversal (Theorem 3).

Theorem 1 says the collisions exist. The next lemma says something stronger and, for the paper’s purposes, more consequential: the collisions are invisible from inside. A bounded agent’s readings of a world larger than its representation space are not merely lossy; they wrap, an element beyond any faithfully represented range being read as if it were an element within one, and no certificate the agent can possess marks a wrapped reading as wrapped.

Lemma 1 

(Wrap: aliasing without internal trace). Let $\mathcal{O}$ be an agent of capacity $(K,H)$ over alphabet A in the structure $\mathcal{M}$, $\left|\mathcal{M}\right|=m$, with $a^K<m$, and let ρ be any total assignment of storage readings to elements of $\mathcal{M}$ realized by the agent.

• (Aliasing)ρ identifies distinct elements; some reading stands for at least $m/a^K\ge 2$ elements, and every element outside a faithfully represented range is read as an alias of one inside it.
• (No internal trace)For every certificate c the agent certifies (writes, step-checks, and retains within budget, per Definition 1), there is a finite structure ${\mathcal{M}}^{\prime }$ with $\left|{\mathcal{M}}^{\prime }\right|\le (r+1)p_{\mathsf{D}}\left(B\right)$, where r is the largest operation arity of σ, $B=min(K,H)$, and $p_{\mathsf{D}}$ is the checking polynomial of $\mathsf{D}$ (the verification also fits inside the run, so $\left|{\mathcal{M}}^{\prime }\right|\le (r+1)H$ as well), in which the same certificate is certified by the same verification steps, is a locally valid derivation, and, for locally sound rules, has a true end sentence, while the readings of every element the verification cites are faithful. Provided $(r+1)p_{\mathsf{D}}\left(B\right)\le a^K$ (true for all capacities beyond a threshold depending only on the signature and on $\mathsf{D}$), possession of c therefore never certifies that a given reading is exact rather than wrapped: the identical certificate arises, with its warrant, in a world where it is exact.

Proof. (1) is pigeonhole. (2) is a trace argument on the verification. Step-checking a record of length at most B takes at most $p_{\mathsf{D}}\left(B\right)$ elementary steps in total (summing the per-step costs over the record), each performing at most one lookup touching at most $r+1$ elements, hence a cited set T of size at most $(r+1)p_{\mathsf{D}}\left(B\right)$. Let ${\mathcal{M}}^{\prime }$ have universe T (padded by one element if empty), with operations agreeing with $\mathcal{M}$ on every cited entry and completed arbitrarily inside T elsewhere. The verification consults only cited entries, so it succeeds in ${\mathcal{M}}^{\prime }$ step for step as it did in $\mathcal{M}$; the record is a locally valid derivation in ${\mathcal{M}}^{\prime }$, and if the rules of $\mathsf{D}$ are locally sound, its end sentence is true there. On a universe of size at most $(r+1)B\le a^K$ the readings of the cited elements can be taken faithful. The certificate is one and the same syntactic object, carrying one and the same warrant, in both worlds; it bears no mark of which world produced it. One caveat delimits the mechanism: a rule with a global side condition (an axiom rule asserting ${\sigma }_{\mathcal{M}}$, say) is checkable only at cost on the order of $m^2$, so a budget-bounded agent cannot deploy it against a large world. Every rule cheap enough to use is local enough to transport; that is why the wrap stays invisible. □

Remark 4 

(The second half of the horizon). Propositions 3 and 4 will say that most truths are out of reach. Lemma 1 says the stronger thing: the world beyond reach does not announce itself as beyond reach. The agent’s picture of the totality is a modular one, wrapped into its own representation space, and the wrap is certificate-invisible. This is the mechanism behind the neutrality argument of Section 7: the same-certificate-different-world move of clause (2) is exactly what blocks any bounded agent from locating its ontology, and it is the static, single-structure form of the indistinguishability result cited there from the finitist literature [16].

Theorem 1 bounds what an agent can hold, and Lemma 1 shows the deficit is silent. Neither yet bounds what an agent can do: a small agent might still, one might hope, certify any given truth on demand, holding at each moment only the certificate at hand. The next section closes that hope.

5. The Horizon

This section carries the paper’s finding. It opens with two counting results, the quantitative face of what survives finitude. It then asks whether the diagonal itself, the engine of the classical theorems, survives, and proves that it does not: a no-compression theorem shows the Gödel template cannot be written over a fixed finite structure, at every scale of its per-symbol coding, and at every scale beyond a structure-dependent constant for arbitrary codings, the residual window being displayed as an open definability problem. The rest of the section takes inventory of what stands in the diagonal’s place: length-horizon facts that are theorems, a complexity remainder that is proved for specific systems and open in general, and one outright reversal, the definability of bounded-fragment truth. Budgeted incompleteness phenomena are the established territory of bounded arithmetic and proof complexity [17,18,19,20]; originality is claimed for the isolation of the no-compression obstruction, for the decomposition it forces, and for the use the assembly is put to in Section 6 and Section 7.

5.1. The Reach Bound

Lemma 2 

(Abundance of truths). Let σ contain the constant 1 and a binary operation +, interpreted in $\mathcal{M}$ so that the value of a closed term depends only on its number of 1-leaves (as in ${\mathcal{M}}_p$, where a term built from 1 by + with ℓ leaves has value $\ell modp$). The closed terms with exactly ℓ leaves are the full binary trees with ℓ leaves, and their number is the Catalan number $C_{\ell -1}\ge 2^{\ell -2}$ for $\ell \ge 2$. Any two such terms $t,t^{\prime }$ yield a true equation $t=t^{\prime }$ of length $O(\ell )$. Hence the true sentences of length at most L number at least $2^{\alpha L}$ for a constant $\alpha >0$ depending only on the syntax.

Proof. 

Terms with the same number of leaves take the same value, so every pair is a true equation: exactly $\left({\displaystyle \genfrac{}{}{0pt}{}{C_{\ell -1}}{2}}\right)$ true equations at leaf count ℓ, each of length linear in ℓ, and $\left({\displaystyle \genfrac{}{}{0pt}{}{C_{\ell -1}}{2}}\right)\ge 2^{\ell }$ for all $\ell \ge 6$ since $C_{\ell -1}\ge 2^{\ell -2}$. The stated $2^{\alpha L}$ bound follows with $\alpha$ absorbing the linear length overhead. □

Proposition 3 

(Vanishing reach). An agent of capacity $(K,H)$ certifies, across all runs of its lifetime taken together, at most $a^{B+1}$ sentences, $B=min(K,H)$: a certificate is a retained record with one designated end sentence, a record is a string over A of length at most B (it fits in storage and is written within the run), and there are at most $a^{B+1}$ such strings. By Lemma 2 the truths of length at most L number at least $2^{\alpha L}$. The certified fraction of length-$\le L$ truths is therefore at most $a^{B+1}/2^{\alpha L}$, which vanishes as L grows beyond a linear multiple of B. The reachable theory of any bounded agent is a vanishing fraction of $\mathrm{Th}\left(\mathcal{M}\right)$.

Proof. 

Immediate from the two counts. Note what is and is not claimed: each individual equation of Lemma 2 is cheap to certify in isolation; the bound concerns the agent’s total reach, not the difficulty of any one target. The lifetime reading is legitimate: certified sentences are counted by distinct records, of which at most $a^{B+1}$ exist as strings, each certifying its one designated end sentence, so sequential reuse of storage generates new instances but never more distinct certified sentences. The totality decides all of $\mathrm{Th}\left(\mathcal{M}\right)$ at grade (ii); the agent, at grade (iii), possesses a sliver. □

Remark 5 

(Possession, not familiarity). The truths counted in Lemma 2 are humble: every one is an instance of associativity and commutativity, and all of them follow from one finite schema an agent might well certify. The bound survives the observation because certification is possession under the retention clause of Definition 1: holding a schema and a derivation procedure is holding two records, not $\left({\displaystyle \genfrac{}{}{0pt}{}{C_{\ell -1}}{2}}\right)$ records, and an instance is possessed exactly when its own certificate is retained. The distinction is the operative notion of bounded knowledge throughout the paper, and it is the classical one in disguise: an effectively axiomatized theory also “holds a schema”; what it provably reaches is still a proper part of what is true. A reader who nonetheless discounts the cohort as trivial variants loses nothing essential: the reach picture is completed by the length horizon and the complexity face of the decomposition (Proposition 4, Remark 9) and does not rest on the cohort alone (see also Objection O2, Section 8).

Remark 6 

(A single hard sentence, system-relatively). The reach bound leaves open whether some single sentence defeats the budget. In the bare evaluation calculus, whose universal rule infers $\forall x\phi \left(x\right)$ from the m premises $\phi \left(c\right)$, one for each element, any derivation of a global clause has length at least m, so every global clause defeats a budget $H<m$ in that system. This is honest but system-relative: a richer $\mathsf{D}$ may derive some global clauses briefly (from ${\sigma }_{\mathcal{M}}$, say). Classically the relativity is removed by the diagonal, uniformly over theories and pointwise in the witness. The next subsection asks whether that route survives finitude, and the answer is no, for a reason that is itself a theorem and is the paper’s central finding.

5.2. The Diagonal Does Not Migrate

Gödel’s construction has a fixed template. A coding of syntax into the domain; a formula $\theta \left(v\right)$ reading properties of coded sentences; a diagonal formula $\delta \left(u\right)$ applying $\theta$ to the self-substitution of u; and the sentence $\lambda =\delta \left(\delta \right)$, which then satisfies $\lambda \leftrightarrow \theta \left(\lambda \right)$ by construction. Over a fixed finite structure the template meets an arithmetic obstruction that no bookkeeping evades.

Lemma 3 

(Mention cost). Let φ be an L-formula and v a tuple of variables, free or quantified in φ. If the truth of φ in $\mathcal{M}$ depends on coordinate i of v (two assignments differing only there give different truth values), then $v_i$ occurs in φ. Consequently a formula whose truth depends on all N coordinates of a tuple has length at least N, and a quantifier block binding an N-tuple contributes at least N symbols.

Proof. 

If $v_i$ does not occur in $\phi$, truth is invariant under reassignment of $v_i$, by induction on $\phi$. Each occurrence costs at least one symbol. □

Theorem 2 

(No compression: the diagonal template is unwritable). Fix the finite signature σ and the alphabet A independently of the coding scale, and require every syntax predicate to be a formula of L or of a fixed finite definitional expansion of L chosen independently of the scale: no scale-dependent primitives (adding a $2N$-ary diagonal relation to the signature changes the structure under discussion, not the theorem). Code strings over A of length at most N as tuples in ${\mathcal{M}}^N$, one coordinate per symbol. Then there is no scale N at which the Gödel template can be instantiated: no formulas $\theta \left(v\right)$, ${\mathrm{Diag}}_N(u,v)$ and sentence $\lambda =\delta \left(\delta \right)$ with $\delta \left(u\right)=\exists v({\mathrm{Diag}}_N(u,v)\wedge \theta \left(v\right))$ exist, where ${\mathrm{Diag}}_N$ defines diagonal substitution on codes and θ is non-degenerate: for each coordinate of its argument there are full-length codes on which its truth depends on that coordinate, as the correctness of any genuine syntax predicate on full-length inputs requires; padding coordinates of shorter strings are exempt (a θ that ignores most of its input can have cheap material fixed points, which carry no self-referential content; Remark 8).

Proof. 

Well-definedness of $\delta \in {\mathcal{M}}^N$ requires $\left|\delta \right|\le N$. But the intended relation ${\mathrm{Diag}}_N$ is injective in u and determines v, so its truth depends on all $2N$ coordinates: $\left|{\mathrm{Diag}}_N\right|\ge 2N$ by Lemma 3; and the block $\exists v$ binds N variables. Hence $\left|\delta \right|\ge 3N+\left|\theta \right|>N$ for every $N\ge 1$. Splitting the scales does not escape: if v ranges over ${\mathcal{M}}^{N_2}$ with $N_2\ge \left|\lambda \right|$ (as it must, for $\lambda$ to lie in $\theta$’s domain), then $\left|\theta \right|\ge N_2\ge \left|\lambda \right|\ge \left|\theta \right|$ plus the diagonal overhead, which is positive. □

Corollary 2 

(Quotation exceeds mention). From Lemma 3 alone: a formula’s arity, hence the number of code coordinates whose values it can genuinely read, is bounded by its length. Consequently no sentence λ contains a subformula whose truth depends on all coordinates of a tuple of arity at least $\left|\lambda \right|$; in particular, under per-symbol coding, or any injective coding of at-least-linear length, no sentence’s truth is sensitive to every symbol of its own code. Not only the standard construction but every self-referential device of this kind fails: self-reading requires reading more than the reader can mention.

Lemma 4 

(Density does not rescue the template). Any injective coding of length-n strings into $\mathcal{M}$-tuples needs $k\ge nloga/logm$ coordinates ($m^k\ge a^n$): a fixed element carries ${log}_{2}m$ bits, so dense coding improves constants, never the linear order. A formula genuinely reading a k-tuple binds k distinct variables, whose names alone cost order $klogk$ symbols once k exceeds the alphabet. For a putative fixed point, $k\ge c\left|\lambda \right|$ then forces $\left|\lambda \right|\ge c\left|\lambda \right|log\left(c\right|\lambda \left|\right)$, a contradiction for all $\left|\lambda \right|$ beyond a threshold $t\left(\mathcal{M}\right)$ depending only on the structure. Hence: under per-symbol coding the template fails at every scale (Theorem 2); under an arbitrary injective coding it fails at every scale beyond $t\left(\mathcal{M}\right)$.

Remark 7 

(The sub-threshold window: an open problem). Below $t\left(\mathcal{M}\right)$ the counting is silent, and the threshold grows with m: solving $c\left|\lambda \right|{log}_2\left(c\left|\lambda \right|\right)=\left|\lambda \right|$ at equality with $c=loga/logm$ gives $t\left(\mathcal{M}\right)\approx ({log}_{2}m/{log}_{2}a)·2^{{log}_{2}m/{log}_{2}a}$, so for $m=2^{100}$ the mention-cost contradiction bites only beyond roughly ${10}^{32}$ symbols, for $m=2^{1000}$ beyond ${10}^{300}$. Whether the window is exploitable is a well-posed question the paper owns rather than hides. Over ${\mathcal{M}}_p$ with p astronomically large, elements below p behave as honest integers under + and ×, and a short, densely coded, Parikh-style diagonal is blocked, if it is blocked, exactly by decoding cost:do short formulas over $({\mathbb{F}}_p,+,\times )$ define the small-integer initial segment and the digit-extraction maps?A short dense-coded diagonal exists if and only if such short decoding formulas exist; the question is adjacent to known hard problems about interval definability in prime fields. The paper’s claims are calibrated accordingly: the operative reading throughout is asymptotic, the structure fixed while budgets and sentences grow without bound, and on that reading the obstruction is total.

Remark 8 

(What Gödel’s construction consumes). The obstruction identifies the resource. A single first-order variable over an unbounded arithmetic carries unboundedly much information: one quantifier absorbs a code of any size at constant syntactic cost, and the numeralwise arithmetic of $\mathsf{Q}$ lets a short formula unpack it. Over a fixed finite structure a variable carries ${log}_{2}m$ bits, and syntax must pay per bit. Gödel’s proof uses infinity not merely as the subject matter of arithmetic but as a syntactic resource: exponential compression of quotation. Equivalently: compression is the license to name magnitudes that never wrap. A fixed finite domain wraps (Lemma 1), the wrap leaves no internal trace, and quotation there can never outrun mention. Withdraw infinity and the diagonal is not defeated but unwritable. Two cautions fix the theorem’s scope. Material biconditionals $\lambda \leftrightarrow \theta \left(\lambda \right)$ are cheap when both sides happen to be true, and carry no self-referential content; the theorem is about the construction, which is what carries diagonal content. And the theorem does not diminish the classical result; it explains what the classical result runs on.

5.3. What stands in the diagonal’s place

Finitude without the diagonal is not finitude without limitation. The inventory has three shelves: counting facts that are theorems, complexity facts that are proved for specific systems and open in general, and one reversal.

Proposition 4 

(Length horizon). A certificate contains its designated end sentence (Definition 1), so no sentence of length exceeding $B=min(K,H)$ is certifiable by an agent of capacity $(K,H)$; and no sentence of length exceeding H has a $\mathsf{D}$-derivation of length at most H in any system. Together with Proposition 3: beyond every budget there are truths, most truths in fact, and the fact is proved by counting alone.

Proposition 5 

(The evaluation ceiling). Drop the local-checkability clause from Definition 1 and admit the one-rule system ${\mathsf{D}}_{\mathrm{tt}}$: assert φ outright, the side condition being truth of φ under exhaustive evaluation against the tables. ${\mathsf{D}}_{\mathrm{tt}}$ is sound, and every true sentence φ has a ${\mathsf{D}}_{\mathrm{tt}}$-derivation of length $\left|\phi \right|$. Under ${\mathsf{D}}_{\mathrm{tt}}$ the truths beyond budget H are exactly those of length exceeding H.

Proof. 

Immediate. The proposition is a ceiling on what any uniform “short truths hard for every sound system” claim could say: without a cost bound on step-checking, none survives. This is why Definition 1 carries the local-checkability clause, and why the two resource measures, derivation length and verification time, must not be conflated: ${\mathsf{D}}_{\mathrm{tt}}$’s steps are short to write and exponentially costly to check. □

Remark 9 

(The complexity remainder). For locally checkable systems the nontrivial question stands: does every sound such $\mathsf{D}$ leave some short truths underivable at every budget? This is the hard-tautologies problem of proof complexity, equivalent in its propositional form to the nonexistence of polynomially bounded proof systems [21]; superpolynomial lower bounds are classical for specific weak systems [19], and the general question is open. Bounded consistency behaves the same way: ${\mathrm{Con}}_{\le H}\left(\mathsf{D}\right)$, the statement that no string of length $\le H$ codes a $\mathsf{D}$-derivation of a fixed absurdity, is true and grade-(ii) decided; for the systems where Pudlák’s bounds are established, $\mathsf{D}$-certificates of length below $H^{\epsilon }$ are ruled out against polynomial upper bounds [20,22]; the full budget-H analogue is open and entangled with the existence of optimal proof systems. Chaitin’s information-theoretic bound gives the same wall a third face [23]. The pattern is uniform: under finitude, all residual difficulty of the first two incompleteness theorems concentrates in proof complexity, proved for specific systems, open in general, and identical on both ontologies, since these are theorems of classical mathematics about finite syntactic objects.

Lemma 5 

(Prefix simulation). Let $\psi (z_1,\dots ,z_d)$ be a formula over $\mathcal{M}$ and $Q_1\dots Q_d$ a quantifier pattern. Then

$$Q_1{z}_1\dots Q_d{z}_d\psi \left(\overline{z}\right)⟺\forall y_1\exists x_1\dots \forall y_d\exists x_d\left[\underset{i:Q_i=\forall }{\bigwedge }{x}_i=y_i\wedge \psi \left(\overline{x}\right)\right].$$

Proof. 

Induction on d; the case $d=0$ is trivial. If $Q_1=\forall$: on the right, $x_1$ is forced equal to $y_1$, so the outer block reads $\forall y_1$ applied to the inner simulation with $x_1:=y_1$; the induction hypothesis applied to $\psi (y_1,z_2,\dots ,z_d)$ makes the inner block equivalent to $Q_2{z}_2\dots Q_d{z}_d\psi (y_1,\overline{z})$, and the whole to $\forall z_1{Q}_2\dots Q_d\psi$. If $Q_1=\exists$: $y_1$ occurs neither in the constraint conjunction nor in the matrix, so $\forall y_1\exists x_1\Phi \left(x_1\right)$ is equivalent to $\exists x_1\Phi \left(x_1\right)$, and the induction hypothesis finishes. The dependency structure is preserved: an existential choice on the right may formally consult the dummy $y_j$ of earlier existential positions, but these are unconstrained and absent from the matrix, so any winning choice function can be taken independent of them. □

Theorem 3 

(Fragment truth is definable: the Tarski reversal). Work over ${\mathcal{M}}_p$ with $p\ge a+1$, in the finite definitional expansion $L_A$ of L by constants naming the alphabet-code elements and the padding element (the expansion is finite, chosen once, independent of n, and conservative for the coding purposes; every finite-structure decidability result is unchanged). Let codes be of sentences in prenex form; prenex is a normalization, not a loss, since every sentence has a prenex equivalent computable at polynomial cost, and stating the theorem for prenex codes settles quantifier polarity structurally. For each n there is an $L_A$-formula ${\mathrm{Truth}}_n\left(v\right)$, v an n-tuple, of length polynomial in n, such that for every $w\in {\mathcal{M}}_p^n$: if w codes a prenex sentence φ of length at most n then ${\mathcal{M}}_p\vDash {\mathrm{Truth}}_n\left(w\right)$ if and only if ${\mathcal{M}}_p\vDash \phi$, and if w codes no such sentence then ${\mathcal{M}}_p\vDash \neg {\mathrm{Truth}}_n\left(w\right)$.

Proof (Proof (construction)). Guess and check, in three layers. Parsing: existentially quantified label variables, a block of fixed size per position, assert a parse structure for the coded string, with position-local consistency clauses against the finite grammar; ill-formed strings admit no consistent labelling. Propositional layer: a guessed evaluation label per matrix position, consistency clauses per connective, the root labelled true; this is circuit evaluation as a certificate. Quantifier layer: let d bound the coded prefix length. The formula carries the uniform real prefix $\forall y_1\exists x_1\dots \forall y_d\exists x_d$, and conjoins the clause $x_i=y_i$ exactly when the coded symbol at prefix position i is the universal quantifier, leaving $x_i$ unconstrained when it is existential and $x_i$ inert when position i is past the coded prefix; correctness of the simulation is Lemma 5, applied with the pattern read off the guessed parse labels. Environments: the matrix labels are tied to the quantified values by linkage clauses, one per atom position. A linkage clause must resolve a data-dependent index, the coded variable index carried at the position, against the fixed variables $x_1,\dots ,x_d$, and variables cannot be indexed by data, so each clause is a case disjunction over the d candidates. The count, displayed once: positions n; prefix length $d\le n$; cases per linkage clause d; name length $O(logn)$; total

$$O(n·d·logn)\subseteq O(n^{2}logn),$$

polynomial in n, and this is where the bulk of the formula’s length lies. The construction is corroborated by exhaustive machine checks at small depths, on the propositional layer and on the pairing layer separately (scripts available from the author); the proof does not depend on them. □

Remark 10 

(The two trivial undefinability facts, and the diagnosis). What remains undefinable is undefinable for reasons of bookkeeping, not of diagonal depth. No single formula defines truth across all lengths, because a formula has fixed arity and codes of longer sentences do not fit its argument places. No formula defines truth for sentences longer than itself, by Lemma 3. Between these two trivialities, Theorem 3 says truth is definable, fragment by fragment, at polynomial overhead: definable from just above, never from within, the shape of the classical Tarski hierarchy with its teeth drawn. Classical undefinability at equal size is thereby diagnosed as another consumer of compression: $\mathbb{N}$ codes its own syntax inside single elements, so the liar can bite; a finite structure spreads codes across tuples, the mention cost intervenes, and the obstruction degenerates into arity bookkeeping while definability, the reverse phenomenon, becomes the theorem. Theorem 1 remains what it always was: a representational bound for agents small relative to the world, an injective-encoding bound, silent about generative or compressed descriptions, which are recipes in the sense of Remark 5.

5.4. The Decomposition Theorem

Theorem 4 

(Decomposition). Over a fixed finite structure, each classical incompleteness-type limitation decomposes into a counting component, which migrates as a theorem, and a complexity component, which is proved for specific systems and open in general; and in every case the classical quotational route, the diagonal, is unavailable, at every scale of per-symbol coding and beyond a structure-dependent threshold for arbitrary injective codings, the residual window an explicit definability problem (Theorem 2, Lemma 4, Remark 7).

Classical limitation	Counting face (theorem)	Complexity face (status)
Gödel I [9,10]	Vanishing reach (Prop. 3); length horizon (Prop. 4)	Short truths hard for every locally checkable system: the hard-tautologies problem, open; proved for specific weak systems (Rem. 9)
Gödel II	${\mathrm{Con}}_{\le H}\left(\mathsf{D}\right)$ true, grade-(ii) decided; its brute certificate is budget-transcendent (formal count $a^{H+1}$)	$H^{\epsilon }$ lower bounds proved (Pudlák); full budget-H analogue open (Rem. 9)
Tarski [14,15]	Whole-structure truth unrepresentable in smaller parts (Thm. 1); arity and mention-cost bookkeeping (Rem. 10)	Reversal: fragment truth definable at polynomial overhead (Thm. 3); no residual difficulty
Speed-up [24]	Reach strictly increasing in B, for systems deriving arbitrarily long true sentences (a reflexivity scheme $t=t$ suffices) (Props. 3, 4)	Resource-hierarchy analogue: per-system separations proved for weak systems, general case open [19,21]

Proof. 

Each cell is the result cited in it; the theorem’s content is the classification. The counting column is unconditional and elementary. The complexity column is classical mathematics about finite syntactic objects, hence ontology-independent, proved and open exactly where proof complexity has proved it or left it open. The quotational clause is Theorem 2. Nothing in the paper leans on more than the displayed status of each cell. □

The decomposition, not a wholesale migration, is the finding. What survives finitude by theorem is the horizon: counting facts about bounded agents in a large world. What remains genuinely hard survives as proof complexity, the same on either ontology. What does not survive at all is the diagonal itself, and its failure to survive is explained by what it consumes: exponential syntax compression, infinity in its role as a compressor rather than as a domain. The classical theorems realize the limitation pattern with compression available; the finite world realizes it with compression withheld, and pays out of counting instead. That the payout is epistemically equivalent for bounded agents is the business of Section 7.

6. Co-extensiveness, Stated Exactly

It is tempting to summarize Section 3 as “incompleteness holds exactly where infinity is interpreted.” The summary needs a clause it usually loses in transmission, and with the clause restored it sharpens into a theorem about where determinacy and effectivity can coexist.

The effectivity clause.

The first theorem requires effective axiomatizability, and the requirement is not decorative. $\mathrm{Th}\left(\mathbb{N}\right)$, true arithmetic, interprets $\mathsf{Q}$, is consistent, and is complete; it escapes Gödel by being effectively presentable by nothing: no algorithm enumerates it, no agent of any capacity possesses it, and its completeness is grade (i) of Definition 2, bivalence by fiat, determinate for no one. The correct co-extensiveness statement is therefore: among consistent, effectively axiomatized theories, interpreting Robinson arithmetic $\mathsf{Q}$ forces incompleteness, indeed $\mathsf{Q}$ is essentially undecidable [11]; declining the interpretation permits completeness, and the finite structures witness the permission (Proposition 2). And the interpretation demands more than unbounded iteration. It demands coding power: enough arithmetic to represent sequences and substitution, the raw material of the diagonal. Presburger arithmetic is the witness that separates the two [25]: unbounded successor iteration, consistent, effectively axiomatized, and nonetheless complete and decidable. Infinity without self-coding yields completeness. What incompleteness tracks is an effective description confronting a domain that codes its own syntax. The finite structures decline self-coding twice over: by cardinality at the scale of the whole (Proposition 1, Remark 2), and by the mention-cost principle at every bounded fragment (Theorem 2), which is why what reappears under finitude is not the diagonal but its decomposition (Theorem 4). Incompleteness is the price of demanding an effective grip on a domain that outruns its describer. Domains that are exhausted by a finite table owe nothing; totalities that are determinate but effectively presentable by nothing pay in a different currency, the currency of grade (i): a completeness no possible agent can cash.

That the finite case is not merely one way of declining the interpretation, but in a precise sense the only way of having determinacy and effectivity together in full, is the content of the following.

Theorem 5 

(Finiteness). Let $\mathcal{M}$ be a structure in a finite signature. The full second-order theory ${\mathrm{Th}}_2\left(\mathcal{M}\right)$, with the second-order quantifiers ranging over all subsets and relations of the domain (full semantics, not Henkin semantics), is decidable if and only if $\mathcal{M}$ is finite. For the forward direction, infinite means: containing a countably infinite subset; the ambient metatheory is classical, and with countable choice every infinite set qualifies.

A version of this characterization is proved as Proposition 9 of [8]. It is restated and proved here to keep the paper self-contained, and because the use made of it, Corollaries 3 and 4, is specific to the present frame.

Proof. (⇐) For finite $\mathcal{M}$ with $\left|\mathcal{M}\right|=m$: a second-order quantifier over k-ary relations ranges over the $2^{m^k}$ subsets of ${\mathcal{M}}^k$, a finite set; exhaustive evaluation, exactly as in Proposition 2 with larger constants, decides every second-order sentence.

(⇒) Suppose $\mathcal{M}$ is infinite in the stated sense, with a countably infinite subset. Dedekind’s categoricity theorem states that the second-order Peano axioms ${\mathrm{PA}}_2$ determine $(\mathbb{N},0,S)$ up to isomorphism under full semantics [26]. For any sentence $\phi$ of second-order arithmetic, consider the second-order sentence over the bare domain of $\mathcal{M}$:

$${\psi }_{\phi }:\exists X\exists R\exists z\left({\mathrm{PA}}_2^{(X,R,z)}\wedge {\phi }^{(X,R,z)}\right),$$

asserting the existence of a subset X, a relation R on X, and an element z of X satisfying the Peano axioms relativized to $(X,R,z)$ together with $\phi$ so relativized. If $\mathcal{M}$ is infinite, some $(X,R,z)$ satisfies ${\mathrm{PA}}_2^{(X,R,z)}$ (build a successor structure on a countably infinite subset), and by categoricity every such $(X,R,z)$ is isomorphic to $(\mathbb{N},S,0)$; hence $\mathcal{M}\vDash {\psi }_{\phi }$ if and only if $\phi$ is true in the standard model. A decision procedure for ${\mathrm{Th}}_2\left(\mathcal{M}\right)$ would therefore decide full second-order arithmetic truth, which is not decidable, indeed not definable at any level of the arithmetical hierarchy [14,26]. This is the paper’s one deliberately classical passage: it evaluates the infinite ontology by that ontology’s own semantics, as an elimination argument must. □

Corollary 3 

(Determinacy meets effectivity only on the finite). Call a structureeffectively determinate in full if its entire theory, second-order included under full semantics, is decidable: every question about the structure, including every question about all of its subsets, has a producible answer. By Theorem 5 the effectively determinate structures are exactly the finite ones. Infinite structures offer a choice: determinacy at grade (i), asserted but presentable by nothing ($\mathrm{Th}\left(\mathbb{N}\right)$, ${\mathrm{Th}}_2$ of anything infinite), or effectivity over a fragment with incompleteness at the fringe (Gödel). Only finitude delivers both, and it delivers them at grade (ii): as finite evaluation objects, not as postulates.

Corollary 4 

(Trichotomy). Effectively described domains fall into three cells. (i) Infinite with self-coding: every consistent effective theory interpreting $\mathsf{Q}$; incomplete, by Gödel. (ii) Infinite without self-coding: Presburger arithmetic and its kin; complete and decidable, the infinitude of the models notwithstanding. (iii) Finite: complete and decidable at the totality (Proposition 2); the diagonal cannot be written, at every scale of per-symbol coding and beyond a structure-dependent threshold for any coding (Theorem 2, Lemma 4, Remark 7), and the limitative facts of bounded parts are counting facts together with proof-complexity problems (Theorem 4). Incompleteness tracks effective self-representation under resource bounds; it does not track cardinality, and the tracking is now explained: cell (i) has syntax compression and pays with the diagonal; cell (ii) has infinity without coding and pays nothing; cell (iii) lacks compression, so the diagonal is unwritable and the horizon does its work.

Cell (ii) is the independent witness for the diagnosis of Section 7: if the phenomenon were about the size of the domain, Presburger arithmetic would exhibit it, and it does not. This corollary and the trichotomy are the load-bearing wall between the mathematics and the philosophy. It converts the vague sense that “a finite world is fully surveyable in principle” into a classification: surveyability-in-principle, read as decidability of the full theory, characterizes finitude. Everything an infinite ontology adds beyond every finite bound is, by the same classification, purchased at one of the two prices just named. Whether the purchase is worth it is the debate; what incompleteness contributes to that debate is the subject of the next section.

7. Neutrality

7.1. The Neutrality Claim

Theorem 6 

(Neutrality of incompleteness, certificate level). Place a bounded agent (Definition 1) in either of two situations: (a) an infinite world of which it holds a consistent, effectively axiomatized, $\mathsf{Q}$-interpreting description T; (b) a sufficiently large finite structure. Every certificate the agent can possess in (a) concerning its incompleteness-type limitations has a counterpart of the same evidential force at the bounded-agent level available in (b), and conversely, the converse read at the level of certificate types: the content of the finite-side certificates, the decomposition theorems included, is classical mathematics about finite objects, derivable by the agent in either situation: a failed proof search recorded to length ℓ; a verified instance family held schematically; a proof-complexity lower bound for a specific system (ontology-independent mathematics, identical in both situations); the conditional metatheorem “if $\mathrm{Con}\left(T\right)$ then $G_T$ is unprovable in T,” possessed as a derivation from an undischarged assumption; and the stronger-system case, a $T^{\prime }$-derivation of $\mathrm{Con}\left(T\right)$ yielding a certificate of $G_T$’s unprovability that is unconditional in $T^{\prime }$, which transfers as the same finite syntactic object with the undischarged assumption relocated to $T^{\prime }$’s soundness, the regress of Objection O6 made explicit. The finite counterparts of the diagonal-involving items are not diagonal objects, which Theorem 2 forbids; they are the same horizon evidence together with the theorem explaining why no diagonal is available, and evidential force, not syntactic shape, is what the neutrality claim requires. What is available in neither situation is the detached, totality-level fact: in (a), certifying “$G_T$ is true and unprovable” outright requires certifying $\mathrm{Con}\left(T\right)$, which the second theorem forbids the agent’s own apparatus; in (b), the corresponding totality-scale facts are grade-(ii) objects beyond budget. Consequently no certificate available to any bounded agent discriminates situation (a) from situation (b).

The proof is an inventory, and the decomposition theorem supplies it. The agent’s possible certificates concerning its limitations divide into the counting face, which the finite ontology proves outright and the infinite ontology exhibits in every finite initial segment; the complexity face, which is classical mathematics about finite syntactic objects and does not mention the ontology; and conditional metatheorems, which are derivations from assumptions and transfer verbatim. The detached facts do not divide, because neither agent has them. The mechanism is Lemma 1(2): every certificate is reproduced, verbatim, in a small world where the agent’s readings are faithful, so no certificate carries a mark of the world that produced it. What requires argument is the philosophical use, so it is argued rather than asserted.

The standard anti-finitist employment of incompleteness runs: the theorems show mathematical truth outruns every effective apparatus; a finite mathematics is an effective apparatus par excellence; therefore mathematical truth outruns finite mathematics; therefore the finite ontology is inadequate to mathematical truth. The employment fails at its first step, and Theorem 6 locates the failure. What the theorems show is that truth outruns every bounded agent’s certificates, and this the finite ontology does not merely accommodate but predicts, by counting (Section 4 and Section 5). The datum “there are truths I cannot certify” is entailed by both hypotheses; a datum entailed by both hypotheses supports neither against the other. For the datum to favour the infinite ontology, one would need the excess truth to be located beyond every finite bound, and no bounded certificate locates it there: every epistemic encounter with incompleteness a bounded agent can actually have, a failed search, an exhausted budget, a lower bound for its own system, is realized inside a sufficiently large finite structure, certificate for certificate. The classical agent’s felt contact with $G_T$ is not an exception, because that contact was never the detached fact: it was the conditional theorem plus faith in $\mathrm{Con}\left(T\right)$, and the conditional theorem transfers. An agent cannot certify from the inside that its wall is the Gödel wall of an infinite world rather than the horizon wall of a large finite one; the two walls are, certificate by certificate, the same wall. The finite ontology does not reproduce classical truth-undecidability, and the argument does not need it to: the decomposition preserves every bounded-agent epistemic consequence of the classical theorems, and that preservation is by itself sufficient to block the anti-finitist inference. This is the decomposition theorem’s evidential face, and it is reinforced by an independent result of the finitist literature: no bounded agent can certify that its universe of objects is infinite at all, since every bounded record consistent with an infinite domain is realized in all sufficiently large finite ones [16]. The trichotomy of Corollary 4 adds the third, independent witness: an infinite domain without self-coding shows no incompleteness at all, so the phenomenon was never registering cardinality in the first place.

7.2. The Diagnosis

If incompleteness is neutral, why has it read so consistently as a discovery about the largeness of mathematical reality? The paper’s diagnosis, offered as a reading and priced as one: because the classical setting hides the agent. In the classical picture the contrast is between a completed infinite totality of truths and an effective theory. The theory looks like a mere instrument, so its defeat looks like the world’s victory: truth exceeds the instrument, hence truth is vast. But the effectivity hypothesis is precisely the agent, abstracted: an effectively axiomatized theory is exactly what a bounded apparatus can wield, the totality of what some idealized-but-bounded certifier could in principle possess. Read so, the classical theorem says: a bounded certifier confronting an inexhaustible domain misses truths. The finite theorems say: a bounded certifier confronting any sufficiently large domain misses truths, and the miss has nothing to do with inexhaustibility in the infinite sense; a domain that exceeds the certifier’s budget suffices. The excess was never information about the world’s cardinality. It was information about the certifier’s finitude, mistaken for cosmology because the classical frame supplied only one thing, the infinite totality, for the excess to be credited to. Undecidability of truth is the shadow the bounded agent casts on an infinite screen; move the screen to finite distance and the shadow falls, with identical shape, on the horizon.

7.3. Consequences for the Standard Employments

Hilbert’s programme.

The received post-mortem holds that Gödel’s second theorem destroyed the programme: finitary mathematics cannot prove the consistency of the infinitary apparatus it was to secure [1,2,27]. The migrated picture adjusts the moral. What the second theorem forecloses is a bounded apparatus certifying itself; that foreclosure holds on the finite ontology too (the ${\mathrm{Con}}_{\le H}$ row of Theorem 4) and so cannot be an argument for the ideal elements’ indispensability or against their instrumental reading. Detlefsen’s instrumentalist defence, that the programme’s core survives if the ideal apparatus is an instrument not obliged to self-certify, receives from the present results a precise underwriting: the ideal, unbounded reading of arithmetic is an idealization of exactly the kind an instrumentalist says it is, and the bounded fragment it idealizes survives on the finite ontology verbatim, instance by instance below the bound. What is lost in the migration is only the pretension the programme never needed: a totality-scale certificate held by a part.

Mechanism and the mind.

The anti-mechanist argument reads the first theorem as showing a human can see the truth of $G_T$ while the machine running T cannot prove it, so the human is no machine [4,5]. The argument is classically defective on grounds long catalogued [28,29]: seeing the truth of $G_T$ requires certifying T’s consistency, which the human cannot do for systems of serious strength. The decomposition adds a structural point. Under finitude there is no Gödel sentence to “see” at all (Theorem 2); what human and machine alike face are length horizons, which treat carbon and silicon identically, and hard instances for their respective systems, which are proof-complexity facts owing nothing to the substrate of the prover. Incompleteness sorts agents by budget, not by substrate; it contains no premise that could distinguish minds from machines.

The platonist moral.

Gödel read the theorems as favouring a mathematical reality exceeding all formal capture. The neutrality theorem does not refute platonism; it withdraws the theorems from its evidence base. A reality that outruns certificates is equally predicted by a large finite world; the phenomenology of inexhaustibility, every attempted step succeeding, every enumeration extended, no experiment revealing an end, is the interior of a bounded agent in a large finite frame, derived rather than denied [8,16]. If there is a case for the infinite ontology, incompleteness is not it.

The residue for the finitist.

Neutrality cuts both ways, and honesty prices the finitist’s side. The finite ontology does not get completeness for its agents (that is the decomposition theorem’s counting column); it does not get a certified frame bound (a bounded part cannot certify the totality’s cardinality, by Theorem 1 applied to that very question); and it must read universal arithmetic claims of ordinary mathematics schematically, instance by instance below the bound, treating the unbounded closure as an idealization. These are real costs, stated in the open. What the finitist gets in exchange is Corollary 3: the only ontology on which determinacy is effectively realized in full, with incompleteness relocated to where it demonstrably operates in everyone’s practice, the certificate boundary of bounded agents.

7.4. Outlook: the Computational Face

The reach gap of Definition 2 has a computational reading that connects it to the central open structure of complexity theory. Checking a retained certificate is cheap by construction; finding one may exhaust any budget. The gap between the two, between verifying and discovering, is on the finite ontology not an anomaly but the expected shape of the horizon: grade-(ii) decisions abound whose grade-(iii) discovery is beyond a given budget while their grade-(iii) verification is within it. The lengths-of-proofs literature founded by Cook and Reckhow [20,21] studies exactly this boundary. The present framework contributes no separation, and claims none; it contributes an interpretation: found-versus-checked is the horizon phenomenon in its computational dress, the complexity face of the decomposition (Remark 9) met as a practice, and its notorious difficulty is what Theorem 1 leads one to expect of questions that quantify over an agent’s whole search space from a vantage no agent occupies.

8. Objections and Replies

O1: “You have merely denied the antecedent.”

The vacuity of Section 3 taken alone would deserve the complaint: noting that finite structures escape Gödel’s hypotheses is elementary and settles nothing. The paper’s content is Section 4 through Section 7: what happens to the phenomenon when its hypotheses fail is itself a body of theorems, the decomposition, including one impossibility (the diagonal is unwritable) that explains the classical construction’s own resources; and the decomposition, not the vacuity, is what defuses the evidential employment. The claim is not “Gödel does not apply, so finitism is safe”; it is “the epistemic content of what Gödel discovered applies everywhere, including inside finitude, so it cannot testify against finitude; and the part that does not apply inside finitude, the diagonal, testifies only to what the classical proof consumes.”

O2: “Completeness of a finite structure is trivial; real arithmetic content is exactly what it lacks.”

Two answers, one for each half. Triviality: the completeness that matters in this paper is grade (ii), a finite decision object, against grade (iii), an agent’s reach, and that contrast is where all epistemic action lives on either ontology; it is anything but trivial, as Section 4 and Section 5 witness, and the possession discipline of Remark 5 is what keeps the contrast honest against deflation by schemas. Missing content: what a finite structure lacks is the unbounded iteration license in the defined sense, the freshness of the successor act, not the act itself (which a cycle supports forever), not induction (Section 2), not the algebra it carries, and not the bounded fragment of arithmetic, which holds verbatim below the bound. The universal closures of ordinary number theory are read schematically, a real and priced cost (Section 7); what the objection cannot claim is that incompleteness is the price’s enforcer, since incompleteness charges both ontologies identically.

O3: “This is folklore: bounded arithmetic, Parikh, Chaitin, finite model theory.”

The components drawn from that literature are classical and cited as such: Gödel’s own speed-up [24], Parikh’s feasibility analyses [17], the finitistic consistency statements of proof complexity [19,22], Chaitin’s information bound [23]. The paper claims originality for none of them; budgeted incompleteness phenomena are the established subject of that literature, and the paper says so where it uses them (Section 5). It claims what the folklore does not contain: the no-compression theorem and its reading of what the diagonal consumes (Theorem 2, Remark 8), the decomposition stated as a classification (Theorem 4), the Tarski reversal in its diagnostic role (Theorem 3), the finiteness classification and trichotomy (Theorem 5, Corollaries 3, 4), and the evidential consequence (Theorem 6), which the folklore, precisely because it lived inside the infinite ontology, never drew.

O4: “The horizon is empirical psychology, not mathematics.”

Definition 1 contains no psychology: K and H are parameters, the results are uniform in them, and every proof is counting. The one place a bound is asserted rather than parametrized is the reflexive one: that an agent cannot certify its own frame’s cardinality is Theorem 1 applied to that question, a prediction of the framework, not an embarrassment to it.

O5: “This is ultrafinitism, and inherits its vagueness.”

No feasibility predicate is used, no sorites is run, and no number is called too large to exist. The frame bound of the finite ontology is a determinate cardinality, not a human threshold; the agent bounds are variables. The position examined here is bounded actualism about the ontology [16], not scepticism about numerals, and the results, being uniform in the bounds, survive any choice of them; the classical treatments of strict finitism’s vagueness problem [3] therefore do not engage the present claims. Yessenin-Volpin’s programme [30] is a different undertaking.

O6: “A stronger agent surveys a weaker one’s wall; does this not revive the anti-mechanist point inside finitude?”

A stronger agent does certify truths beyond a weaker one’s horizon; that is the hierarchy row of Theorem 4, and it holds symmetrically for stronger machines surveying weaker humans. What would revive the anti-mechanist argument is an agent surveying its own wall from above, and that is what nothing possesses: under finitude there is not even a Gödel sentence to “see” (Theorem 2), and a totality-scale certificate is what Theorem 1 denies to every part, human or mechanical.

O7: “Why should conclusions about mathematics follow from one toy structure?”

Every theorem is stated for arbitrary finite structures under displayed hypotheses; ${\mathcal{M}}_p$ fixes ideas and magnitudes only. Nothing turns on modular arithmetic. The generality is the point: migration is not a feature of a cleverly chosen example but of boundedness as such.

O8: “Grade (ii) itself violates the finite ontology: the evaluation object of a deep sentence has size $m^q$, beyond any frame bound.”

The objection confuses the existence claim with the object. Grade-(ii) status is certified by a computation on a recipe: the evaluation object is specified (a tree of quantifier instantiations over the structure’s tables), its size is computed exactly, and the specification and the size computation are short, budget-level records. Nothing in Definition 2 requires the object to be realized, any more than the candidate count $a^{H+1}$ in the Gödel II row of Theorem 4 is realized by a search; the paper treats such magnitudes as formal counts throughout, and says so at each occurrence. What distinguishes grade (ii) from grade (i) survives this reading intact: for grade (ii) a finite decision procedure over the structure’s tables is specified and its cost computed; for grade (i) no procedure of any specification exists, which is exactly the currency in which $\mathrm{Th}\left(\mathbb{N}\right)$ pays. A finite ontology owes existence to its domain, not to every combinatorial object specifiable over its domain; the distinction between the grades is itself schematic and certificate-level, which is what such an ontology demands of its own bookkeeping.

9. Concluding Remark

The incompleteness theorems are permanent mathematics; the question was never their truth but their address. Read at the address they were posted to, an infinite domain under effective description, they seem to reveal a world too large for any grasp. Delivered to a finite address they say the same words with a different referent: any world is too large for a bounded part’s grasp, provided only that it exceeds the part’s budget, and the excess registers identically whether the world outruns every bound or merely the part’s own. A finite world is complete, and decides everything; each of its parts certifies almost nothing of it, cannot hold its image, cannot cheaply vouch for its own apparatus, and meets truths beyond every budget it will ever have. What such a world lacks is not limitation but one device: the diagonal, which cannot be written there, because Gödel’s construction spends a currency, exponential compression of quotation, that only an infinite domain mints. The limitation stands without the device. That is incompleteness enough, and it is exactly the incompleteness we in fact observe. What the theorems measure is the knower. About the size of the known, they are silent.