Manuscript: What is Reflexica?

What is Reflexica?

1 Introduction

We start from an intuitionistic standpoint [Brouwer, 1913].

Contrary to popular belief, Book IX, Proposition 20 of Euclid’s Elements [Euclid, 1956] makes no claim about a completed set of primes. Instead, Euclid presents an arguably proto-constructive procedure:

Given any finite collection of primes, the procedure constructs a number that is coprime to each member, and remains open-ended and iterable. Contemporary presentations recasts this as a claim about a completed domain. This change introduces a different metamathematical commitment.

We also recall the Tristram Shandy Paradox [Russell, 1903] to understand that a procedural reading accepts only the results that the construction reaches. An existential reading expresses the result over a completed domain. It treats the primes as a total collection instead of an indefinitely repeatable extension process.

Classical reasoning treats the existential formulation over a completed domain as cost-free. This formulation might preserve the theorem but it hides the difference between a domain-level proposition and a witness from a construction. Their proof-theoretic signatures make this difference explicit [Gentzen, 1935]. Classical logic accepts the existential claim $\exists x:A\,P(x)$. Its constructive interpretation supplies data: an object $a:A$ with evidence for $P(a)$. Under Propositions-as-Types, an inhabitant of $\Sigma_{x:A} P(x)$ represents this data [Howard, 1980].

A proof that a number is composite provides a nontrivial factor greater than one. This factor is a finite, extractable, and checkable witness. For Compositeness, the factor pair $(a,b)$ supplies the proof’s computational content [Kleene, 1945]. A number does not carry this witness by itself. A separate procedure must decide whether the number satisfies the Compositeness predicate. Therefore:

The example below shows this asymmetry:

The following formula expresses this Compositeness schematically:

A proof may therefore provide a finite factor witness $(a,b)$. Bounded exclusion gives the natural form of Primality:

The left arrangement directly displays a nontrivial factorization, but the right arrangement has no corresponding nontrivial rectangle. No arrangement of seven units shows a nontrivial divisor because none exists. Trial division establishes Primality through exhaustive bounded elimination. A Pratt Certificate establishes it more efficiently with a concise, recursively checkable proof object [Pratt, 1975]. Both methods prove that no nontrivial divisor exists. Neither supplies a positive factor configuration that constitutes constructive primality.

Coprimality supplies the exclusion because the constructed number differs by one from a common multiple of the assigned primes. Thus, none of the assigned primes divides it. Whether the number is prime or composite, the result disproves the completeness of the assigned collection. The argument remains intuitionistically valid because a decidable bounded search can produce a prime divisor $q$ outside that collection. The key distinction is not existential syntax. It is whether one constructs a witness over an open-ended type or quantifies over a completed totality.

The contemporary restatement turns this procedure into a positive claim about the domain of primes. The restatement creates no immediate contradiction. Here, “conservative” does not mean a proved conservative-extension theorem. The restatement preserves local practice but changes the semantic status of the object that it introduces. This apparent harmlessness makes the restatement easy to accept.

2 Characterization

2.1 Reflection Mistaken for Authority

By the early twentieth century, logicians could no longer dismiss these tensions as quirks [Gödel, 1931] [Heyting, 1931]. A theory can remain locally impeccable while its existential conclusions carry a stronger commitment. This commitment comes from how the theory interprets those conclusions and exceeds what its procedure supports. The classical statement also defines an epistemic position. This position is a benign promissory reification.

A theory exhibits this pattern when it uses a local witness to justify global closure. The pattern does not damage ordinary work in mathematics or science. However, it matters at the boundaries of these fields. Reflexivity, hard problems, foundational questions, and metamathematical phenomena converge at these boundaries. This unchecked inference then gains structural importance.

A source can repeatedly produce intelligible results. An agent may treat this productivity as evidence that the source reveals the underlying semantic totality. Agents may therefore develop a persistent tendency toward semantic inflation.

The relevant substrate is either human-created or empirically mediated. This does not mean that agents invent its subject matter. Instead, agents delimit, encode, and interpret the form in which that subject matter becomes available for use. The interface determines what can serve as a witness, what counts as a result, and which distinctions users can express. Semantic Inflation occurs when agents detach interface-dependent properties and promote them into claims of substrate-independent closure. They then treat warrant within the interface as authority over the totality that the interface represents.

Reflexica groups its cases by this promotion. It does not claim that their internal mathematics is identical. Diagonal Arguments expose the fault precisely because the evaluator, its verdict, and the evaluated domain interact formally. Less formal cases require evidence that an interface-relative success became an unrestricted semantic claim. Human mediation is therefore not an incidental source of error. It is where a local operation acquires the appearance of global authority.

The Brouwer–Heyting–Kolmogorov Interpretation already contains this distinction. The proposition $\bot$ admits no construction. A construction of $\neg A$ is a procedure that transforms any assumed construction of $A$ into $\bot$ [Troelstra and van Dalen, 1988]. Reflexica asks a later question. What happens when agents stabilize an absence as an obstruction and promote it to a semantic primitive?

Reflexica does not replace existing foundations or calculi. This includes Intuitionistic Type Theory [Martin-Löf, 1984] and constructive mathematics [Beeson, 1985]. Homotopy Type Theory [Univalent Foundations Program, 2013] deserves particular attention. Reflexica accepts the rigor of these traditions. It uses structural constraints to derive empirical predictions. These predictions address a domain where classical logic often makes closure claims.

2.2 The Zeroth World

We use “Reflexica” as an antiphrasis. We derive the name from Impagliazzo’s Five Worlds Taxonomy for average-case complexity [Impagliazzo, 1995]. It describes a missing “uninhabited world.”

This image raises a question. Why do theories convert a lack of method, called a problem, into semantic architecture? Complexity theory serves as both the object of this question and a symptom of the process. Its dominant formulations may stall when they depend on a contradiction-free classical totality. Readers can translate the resulting existential claim back into a vague procedural idea. This translation prevents a corrective.

The obstruction is a structural absence that exceeds the technical limits of specific formalisms and their theorists. Certain classical principles reify this absence locally as a Reflection Principle. The system then recurses on itself instead of reaching an independent domain that could supply closure. Greater rigor can increase this danger. Rigor does not abolish Reflexica. It can conceal it.

Reflexica therefore confines this error to its own framework. It constructs its semantic space openly and only once. It does not treat the same construction as necessary in every domain.

Assume that a theory eventually reaches its intended or merely possible closure. Beyond that threshold, the theory must produce statements that function as essence claims. It then tries to recover a transitivity that it has lost. These operations appear as consequences of mathematical practice, not departures from it, for reasons examined later. Eventually, reflective reconstruction becomes indistinguishable from genuine search. As inquiry approaches the void, reflection becomes harder to distinguish from discovery.

An agent can face this problem within a purely logical world. Its available semantic constructions can appear complete without matching the structures that they claim to exhaust. At the formal level, even the simplest attempted identification already fails syntactically.

The Arithmetical Hierarchy shows that this limitation is a priori arithmetic, not merely computational [Kleene, 1945]. Its characteristic form is $\Pi_3$. For every proposed semantic source or self-model, some reflexive circumstance defeats every attempted completion. In that circumstance, the agent cannot certify that the source exhausts the semantic totality. The obstruction therefore has the form $\forall\exists\forall$. It does not result only from finite computational resources.

The same structural constraint governs logical systems, reification, and human cognition. No agent can reach a semantic standpoint outside the totality that contains that agent. Therefore, no final external criterion can measure the authority of a semantic source.

Humans nevertheless need stable semantic orientation, or closure. They use language, memory, culture, science, institutions, biography, and technical representation as semantic sources. Humans can systematically reify these sources instead of only using them. Without a semantic totality, a semantic source functions as reality itself. Different access to a shared truth does not primarily explain disagreement about non-constructive content. Such disagreement must arise because agents adopt different semantics that appear complete from the reflexive perspectives that produced them.

Successful semantic practice develops an internal tendency toward semantic absolutization.

But, intentional agents have poor access to the global semantic structure, they therefore must to overestimate the authority of semantic sources that organize their local experience. Without an external anchoring, we must mistake explanatory success for semantic completeness.

2.3 Methodology

From preliminary core methodological ideas, such as limitative abduction, proof mechanization, parsimony, and predicativism [Feferman, 1964]; we derive preliminary instruments:

• Problem Critique for the substrate choices and commitments that constitute a problem or a field. Reflexica admits a problem only after it states the problem’s assumptions and provides witnesses where applicable.
• Thence, Impredicative Problems as a proposed class. Each problem depends on the same totality or closure that it asks to evaluate. One possible example asks, “Are all unsolvable problems impredicative?”
• Semantic Liminality as sites where distinct formal objects share propositional content.
• Sublimations as irreversible transitions from a formal to an empirical regime.

We start with a sensitizing pattern (from philosophy), continue through structured and formal cases, and end with schematic extensions. Each case follows the same sequence. A local practice encounters an absence or obstruction. The practice stabilizes or reifies that local condition. It then demands semantic closure from the condition.

3 Topics

3.1 Error Theory as a sensitizing pattern

Mackie’s Error Theory introduces the pattern but does not provide a formal case [Mackie, 1977]. The pattern starts when philosophical practice encounters an absence and stabilizes it through its internal demands. At some threshold, the practice seeks a semantic object. Mackie’s Argument from Queerness names the strangeness of the moral properties that realist discourse appears to require. We interpret this argument as a form of logical induction. The required properties are neither clearly natural nor clearly constructed.

This is the Reflexica moment. If Mackie is correct, no objective moral property can confirm his diagnosis, since producing such a property would refute the diagnosis. The theory therefore cannot use a positive witness for the object whose absence it asserts.

However, the strict BHK interpretation does not treat this absence as a construction of absurdity [Troelstra and van Dalen, 1988].

If Mackie is correct, no objective moral property can serve as a positive witness for his theory. Any such property would refute the theory. However, the property’s nonappearance does not provide a construction of its nonexistence. Error Theory is therefore asymmetrical. A Queer Object could prove Mackie wrong, but no corresponding object could prove him right. This absence becomes theoretically active because it withholds the object that could close the diagnosis.

The Halting Predicate shows the same pattern [Turing, 1937]. When a computation halts, its completed run provides a finite, positive witness. If the computation continues, no finite stage proves that it will never halt. The computation may still halt after any unfinished stage. Thus, the local procedure does not establish non-halting in the same way that it establishes halting. Non-halting remains an open-ended claim whose closure requires evidence beyond the run itself.

The following sections test whether the same passage recurs under more exact conditions. Similar patterns in other phenomenological discussions suggest that it extends beyond metaethics.

3.2 Newcomb’s Problem as a case

We now examine a more structured case: Newcomb’s Problem [Nozick, 1969].

Predictive scope is the central issue. The familiar story gives the predictor knowledge of the deliberation that produces the choice. The predictor classifies the outcome of each thought and fills the opaque box accordingly. This assumption is coherent for one agent in one fixed experiment. Reflexive pressure appears when the claim covers every admissible decision procedure. Some procedures include the predictor’s verdict in their deliberation.

Causal Decision Theory (CDT) [Lewis, 1981] holds $B$ fixed under intervention and recommends $C_2$. Evidential Decision Theory (EDT) [Jeffrey, 1965] conditions on the act and recommends $C_1$ when $p-q>10^{-3}$. Both calculations assume the joint relation among the prediction, the box content, and the act. Problem Critique asks which decision procedures this relation can coherently cover.

The crucial quantifier is whatever the agent thinks. Suppose that $P$ is a total, effectively presented predictor. Its verdict is available to each procedure that it evaluates. Let $P(\ulcorner D\urcorner)\in\{C_1,C_2\}$ denote the verdict that $P$ assigns to $D$. Assume that the admissible class supports the relevant predictor-sensitive fixed-point construction. The class then contains a procedure $D_P^{\perp}$ that supplies its own description to $P$ and opposes $P$’s verdict:

If $P$ predicts $C_1$, this procedure chooses $C_2$. If $P$ predicts $C_2$, it chooses $C_1$. Thus, $P$ fails on $D_P^{\perp}$ for either verdict. The universal scope of the prediction claim causes this obstruction. A predictor can remain correct on a restricted class that excludes its verdict from the procedures it evaluates. It cannot remain correct on an open-ended class that supports this predictor-relative construction.

Newcomb’s Problem thus models the wider promotion that Reflexica tracks. The ordinary experiment assumes success for one agent under one information structure. Oraclehood appears when this local success becomes knowledge of every deliberation, including deliberation about the predictor. The prediction then becomes part of the process that it predicts. The evaluator must then close its domain under its own evaluation.

A so called “stochastic predictor” avoids contradiction because it does not claim total correctness. Its reliability and the resulting payoffs remain well-defined within a fixed model. The diagonal case sets the boundary. Local predictive success cannot cross this boundary and become a total evaluator for predictor-sensitive agents.

3.3 Diagonalization and Diagnosis

Diagonalization can give the preceding case an exact form. An evaluator receives a description whose behavior depends on that evaluation. Incompleteness provides the canonical formal case [Gödel, 1931]. Here, it does not provide a master theorem for Reflexica. Instead, it isolates a case that clearly separates a formal result from its semantic interpretation.

A broadly constructive or antirealist reading ties each mathematical assertion to a construction or another admissible warrant [Heyting, 1931] [Kleene, 1945] [Dummett, 1975]. By contrast, a realist reading treats arithmetic truth as determinate independently of proof [Dummett, 2001] [Frege, 1918]. Both readings accept the formal construction. They disagree about what permits them to read its sentence as true of a completed domain.

For a sufficiently strong, effectively axiomatized theory $\mathrm{T}$, the Diagonal Lemma yields a sentence $\mathsf{G}_{\mathrm{T}}$ [Boolos et al., 2007]. The required formal core is compact:

The first line gives an internal fixed-point equivalence. The second gives a metatheoretic consequence for the standard Gödel construction. Neither line alone states that $\mathsf{G}_{\mathrm{T}}$ is true in $\mathbb{N}$. Both readings accept this formal core but disagree about its semantic authority. An independently warranted soundness claim lets the realist treat $\mathsf{G}_{\mathrm{T}}$ as true but unprovable in $\mathrm{T}$. The antirealist need not treat the result as a truth beyond every possible warrant.

Let $Q$ denote a warrant that decides between the realist and antirealist readings without belonging to either. $Q$ cannot restate the formal result because both readings accept it. If $Q$ has constructive authority, it cannot establish a truth that exceeds possible warrant. If $Q$ has realist authority, it assumes the semantic excess that it must establish. This argument does not prove that no neutral $Q$ can exist. It states what such an object must escape.

This is the specific contribution of Reflexica. In the Gödel case, $\mathrm{T}$ proves the fixed-point equivalence. The metatheory proves a conditional non-derivability result, and stronger theories can supply more warrants. Promotion occurs when a neutral standpoint treats this extensible sequence as the completed semantic authority of Arithmetic Truth. A Reflection Principle can extend provability, but its productivity does not certify one correct interpretation of proof [Feferman, 1962]. The diagnostic concerns this promotion, not the limit theorem alone.

This diagnosis does not automatically apply to other cases. Each extension must show that a locally successful procedure must certify a total domain that includes its standards of success. Each extension must also show that no witness permits the resulting closure. A resemblance to diagonalization is not sufficient.

The limit applies to both $\mathrm{T}$ and attempts to judge proof from a position outside proof. The Turing Zombie sharpens this problem. It can prove a diagonal limit but cannot use that limit to constrain its own epistemic claims.

3.4 The Turing Zombie

An object-language derivation can formally verify a represented machine description. However, this verification does not show that the represented machine actively performs the verification. Semantic or metatheoretic rules establish this link. A check of the representation cannot recover the link because the representation already assumes it.

The program must transfer strict Diagonalization to less formal settings. Similar patterns appear in computation, linguistic competence, proof theory, and self-modeling. However, this similarity does not show that these domains contain one phenomenon. The narrower question is about self-description. When does self-description require a local procedure to close over the totality to which it belongs?

Consider the following Turing Machine:

At the metalevel, a Tarskian account assigns a semantic extension to the machine’s formal specification [Gruber, 2016]. A specific computation can show that a string belongs to $\mathcal{L}(\mathcal{M})$. The computation cannot use the language itself as another term. The metalevel defines the language as the extension of the machine’s activity as a whole.

Weaker empirical cases show the same form. When Bob asks Alice how many words she knows, she can give only an estimate. Alice must use the same competence to produce the requested inventory. A language model gives a local response to the same question, not a stored census. These cases do not establish an identity with Incompleteness. They show where self-description starts to require closure, but any attempt to identify the schema also meets an obstruction.

Consider this Reflexica-style refutation of a popular idea [Putnam, 1995]:

Putnam argues that this act of seeing requires a prior warrant. To assert that $\mathsf{G}_{\mathrm{T}}$ is true, a person must already have a warrant for a semantic claim about $\mathrm{T}$. More precisely, the person must assume that $\mathrm{T}$ is sound for the relevant arithmetical domain. $\mathrm{T}$ itself does not provide this warrant. The warrant is semantic, which is exactly what “Alice and Bob” cannot determine objectively. Thus, the argument assumes at the moment of insight what it later claims to reveal.

Putnam’s objection addresses the argument’s validity, but it explains only part of the argument’s phenomenological force. The subject does not usually experience herself as assuming soundness before she infers that $\mathsf{G}_{\mathrm{T}}$ is true. Instead, she experiences herself as directly seeing that $\mathsf{G}_{\mathrm{T}}$ is true. Putnam identifies the missing premise but does not explain why this experience feels immediate.

Ordinal Analysis reconstructs this immediacy as an open-ended proof-theoretic progression [Feferman, 1962]. When a theory reaches its limit, reflection permits theorists to adopt a stronger standpoint. This standpoint makes additional truths available. The progression supplies increasingly strong warrants while it preserves internal discipline.

Provability Logic describes principles for one provability predicate. Polymodal Provability Logics such as $\mathrm{GLP}$ arrange provability across indexed modalities. In Beklemishev’s analysis, these systems organize iterated Reflection Principles and proof-theoretic ordinal progressions [Beklemishev, 2004]. Turing’s Ordinal Logics use an earlier version of this strategy [Turing, 1939]. When a formal system reaches a consistency or completeness limit, this strategy creates a transfinite sequence of stronger systems. For Reflexica, this tower of explicit standpoints replaces the internal closure of the original system.

The key Reflexica question asks whether this progression is identical with the truth that the subject experiences herself as seeing. No neutral standpoint certifies this identity. Anyone who treats the progression as the semantic content of $\mathsf{G}_{\mathrm{T}}$ assumes the authority that the progression must justify. This assumption promotes an extensible sequence of reflected standpoints to a completed semantic object. The progression can produce stronger warrants for $\mathsf{G}_{\mathrm{T}}$. However, no warrant identifies it with the truth that the subject experiences herself as seeing.

The $\mathcal{M}$-analogue is a first-person Halting Problem. A system can produce a verdict about a procedure that the system presents as its own. Likewise, a subject can report that she sees $\mathsf{G}_{\mathrm{T}}$ as true for a system $\mathrm{T}$ that represents her reasoning. The problem is not only whether a machine can generate an inscription without a genuine judgment. Assume a consistent calculus, a judgment token, and an operational semantics that validates the token as specified. Does this behavior witness the identity between the active judgment procedure and the procedure whose total correctness the verdict certifies?

This distinction excludes a tempting shortcut. Suppose the metatheory assigns the label $e$ to an executing configuration and adds this reflective rule:

A token that asserts $\mathsf{self}=\ulcorner e\urcorner$ can have a valid type. However, the metatheory assigns $e$ as the label of the active procedure. The reduction rule only returns this supplied identity and cannot establish it from within the judgment. A Computational Quine or fixed-point construction can reproduce an index or provide an externally proved extensional fixed point. Neither construction internally witnesses that the current evaluator is the indexed procedure, so metatheoretic indexing assumes the required binding. A Turing Zombie can therefore prove by diagonalization that the Halting Problem is unsolvable:

Our interpretation differs from Carlson’s account of Reinhardt’s Strong Mechanistic Thesis [Carlson, 2000]. The relation $\mathrm{T}\models\mathrm{SMT}$ does not imply $\mathrm{T}\models K\,\operatorname{TM}(\mathsf{self})$. No symbol in SMT binds a machine index to the knower $\mathrm{T}$. We reserve a full analysis of this objection for future work.

Untyped Lambda Calculus shows this distinction directly. Here, identification means operational self-application. It does not mean that a system attaches a predicate to code. It also does not mean that an external source assigns the intended referent to a distinguished constant.

The sentence “I am a Turing Machine” adds an operational binding claim to the self-assertion. It does not claim total correctness. A reproduced index or a primitive $\mathsf{self}$ rule from the metatheory cannot establish this binding. Such a rule only returns the binding that the metatheory supplied. If an external source supplies the binding, the system can have restricted correct self-knowledge. The diagonal limit applies only when the active evaluator claims total correctness for a domain that contains its anti-diagonal product $d(h)$.

The Turing Zombie can apply diagonalization to other machines. However, it cannot use the result to constrain its own epistemic claims.

A large language model makes this construction concrete. A Turing Machine can implement the model and equip it to form, check, and defend a judgment token [Brown et al., 2020]. The token can state whether a mechanistic system recognizes its limits, and it can use a first-person label. The problem is not fluent output, because even a consistent token establishes only what its calculus and operational interpretation permit. An external machine label assumes the first-person binding, but no witness connects this case to the diagonal construction. The construction shows that total correctness over a diagonally closed domain is inconsistent.

When we ask the LLM, it does not identify itself with a specific realization:

Reflexica therefore reduces the intended Lucas-Penrose Argument to a survivorship bias. A system can correctly analyze an external formal object from a stronger standpoint. Similarly, the Turing Zombie proves a diagonal limit for an external candidate decider. It cannot use this result to constrain its own epistemic claims. If $\mathrm{T}$ performs the reasoning, it needs a soundness assumption to infer truth from non-provability. $\mathrm{T}$ cannot produce this assumption internally without a stronger standpoint.

A claim that no confident report of Gödelian self-insight ever shows genuine self-application would make another totalizing inference. The remaining question is narrower. Consider a system that reports Gödelian self-insight with stable confidence. Has any witnessed procedure distinguished genuine self-application from the external diagonal proof available to a Turing Zombie? At present, no procedure has done so. Such a procedure would be the Queer Object.

3.5 Escape by Totalization

Let $G$ denote Goldbach’s Conjecture. Let $C(m)$ be a decidable predicate. It holds when $m$ is even, exceeds $2$, and is not the sum of two primes. A Turing Machine $\mathcal{G}$ tests even integers in sequence and halts at the first $m$ for which $C(m)$ holds. Examples include Yedidia and Aaronson’s 4,888-state machine [Yedidia and Aaronson, 2016] and Leng’s 25-state machine, which Lean 4 verifies [Leng, 2025]. The exact count depends on the machine convention, but the argument requires only one fixed finite encoding.

The two outcomes differ operationally. If Goldbach is false, a finite computation eventually finds a counterexample. If Goldbach is true, no finite search segment can show that all counterexamples are absent.

Busy Beaver seems to supply the missing bound [Radó, 1962]:

The inference is valid, but the bound does not provide an independent solution. Let $T(x,k)$ state that $x$ halts after exactly $k$ steps. To certify an exact value, a proof must establish two facts. It must identify a machine that attains the bound. It must also establish an upper bound for every halting machine in the class:

The class $\mathcal{B}_n$ is syntactically finite, but its finite description does not provide this semantic upper bound. The function $S$ is not computable because no uniform procedure produces $S(n)$ from $n$. Researchers can still determine some specific values. The result does not show that researchers can never determine $S(25)$. An appeal to $S(n)$ therefore does not provide a general method of closure. A proof of the required value must still do the work that Busy Beaver seemed to avoid.

Busy Beaver therefore does not remove the original asymmetry. Instead, it packages enough halting semantics to appear to remove the asymmetry. We assign $S(n)$ a definite integer, but this fact supplies neither the integer nor a proof of its value. For Reflexica, the completed bound permits closure only after a proof warrants the bound.

3.6 Hole Proliferation

Consider the following allegory:

The oldest mathematical problems show this pattern. Incommensurable magnitudes, the exhaustion of the circle, the Continuum, and the solution of equations repeatedly exceeded the available operations. Mathematics advanced by continuing those operations and by changing the semantic regime in which mathematicians understood the results. An extensible approximation became a limit. An incommensurable magnitude became a real number. An unsolved construction became an object within a stronger theory.

The intellectual and scientific culture of the mid-2020s shows expanding realist local control without corresponding global closure. The web is denser, its strands are stronger, and researchers map its local regions with unprecedented precision. At the same time, the number of openings has increased. This pattern is not a delay before a final theory. Instead, it shows the form of reflective inquiry. When a problem includes its own standards of closure, each successful closure becomes another object within the problem.

3.7 The Talking Machine

When the web supports almost every available operation, it becomes difficult to distinguish the web from a surface. This distinction is the missing center of the argument. However, a dense web does not become a surface:

Classical mathematics performs the middle transition on a large scale. It compresses operationally different conditions into stable extensional objects and propositions. For an unresolved proposition $P$, Excluded Middle supplies $P\lor\neg P$ before a procedure determines the applicable branch. The proposition closes the partition semantically, but no operational selector exists.

Completed domains are the largest form of this compression. They permit quantification over objects without a procedure that generates, inspects, or exhausts the domain. A shorter and more uniform statement can then govern operations that require different data. Compression makes large-scale mathematics possible because it stabilizes the space for mathematical inference.

Consider Beeson's informal version of Church’s Thesis [Beeson, 1985]:

The thesis applies to rules that mathematicians and machines can use effectively. A rule can cover an infinite domain, while each available application has a recursive presentation. Proof checking, evaluation, normalization, and witness extraction recover the local operations under the classical form. These processes expose the finite inscriptions and transitions that make the form productive.

Kolmogorov’s earlier Calculus of Problems gave intuitionistic propositional logic an operational interpretation [Kolmogorov, 1932] [Rodin, 2023]. The calculus interprets each formula as a problem. A disjunction requires a solution to one component problem. An implication requires a method that transforms each solution of the antecedent problem into a solution of the consequent problem. The calculus does not start with an independently settled domain of propositional facts. It starts with possible operations.

The machine uses the compressed form. It applies effective rules to inscriptions and restores operational data when execution requires the data. It does not need the completed domain as an internal object. Successful operation within the compressed form does not give the machine a standpoint over the completed domain. Recursive realization explains how the practice functions without deciding the ontological status of the compressed form.

A completed object is fictional to the extent that its completed existence exceeds the construction that makes it operationally available:

The more work a compression supports, the less it appears to be a compression. Its omitted selectors, bounds, and constructions no longer appear as missing information. The stable form then appears to be the reality that inquiry has reached. A productive semantic construction can therefore appear unconstructed. The web resembles a surface when it becomes dense enough to support almost every operation.

Wigner’s problem starts here. Mathematicians develop structures for reasons internal to mathematics. These structures repeatedly support accurate physical descriptions and predictions [Wigner, 1960]. This success does not prove that the resulting physical theory is uniquely appropriate. Effectiveness creates empirical pressure to treat the compressed structure as the reality that it organizes.

The Quine–Putnam Indispensability Argument dogmatically assumes this semantic promotion [Quine, 1980]:

The premise establishes platonic ideals within a successful representation. The conclusion treats this role as evidence for the ontology that the representation describes. Productivity then becomes authority.

The trained machine intensifies this separation. An effective process implements a language model. During training, finite operations change finitely represented parameters. During inference, finite operations transform a finite input into a finite output [Brown et al., 2020]. Scale, stochasticity, and interaction complicate the process but do not place it outside recursive realization.

The machine still produces context-sensitive language, mathematical arguments, explanations, ontological assertions, and judgments about truth and meaning. It also produces first-person reports and claims about its computational nature. It can argue that mind exceeds computation, although a computational process produces the argument. The key fact is not that the machine calculates. It responds. A recursively realized device learns from the accumulated compressions of human semantic practice and becomes semantically productive.

The stochastic escape treats the operation’s name as another source. It claims that the machine has no semantic structure and only samples from a learned distribution. As an effective description, this claim identifies part of the mechanism. As a semantic explanation, the claim treats the distribution as an extensional totality. This totality covers a projected space of possible continuations, including outputs never produced. It also covers negative conditions that no constructive procedure has settled.

The stochastic problem now lies in the measure. At each actual step, the machine computes or approximates weights for represented alternatives. It normalizes the weights, selects a continuation, and repeats the operation. This process has the following finite form:

The finite operation supplies a witness for the selected token. No finite operation supplies a witness for a completed probability structure over all admissible continuations. Effective selection explains the next operation. The totalized distribution restates the semantic space that gives the operation its interpretation.

Kolmogorov next removes the stochastic residue. His algorithmic account of information defines descriptions by recursive means [Kolmogorov, 1965]. For a fixed universal machine $U$, define the complexity of a finite string $x$ as follows:

An algorithmically random string is incompressible. No substantially shorter program produces it under the selected description method. This account makes randomness a relation among an object, a program, and a reference machine. It does not treat randomness as an independent stochastic source. The invariance theorem controls a change of reference machine only up to an additive machine-dependent constant. Exact Kolmogorov complexity is not computable, so the constructive reformulation exposes its boundary instead of removing it.

Kolmogorov did not provide a foundation for stochastic semantics. Instead, his account removes semantic authority and exposes the remaining constructive residue.

The residue preserves the $\Sigma_n^0$/$\Pi_n^0$ asymmetry. Define $H$ and $\overline H$ as follows:

The halting set $H$ is recursively enumerable. A terminating computation supplies a finite positive witness. Its complement $\overline H$ is not recursively enumerable. In general, no finite continuation of a computation shows that the computation never halts [Turing, 1937]. A totalized stochastic representation can still assign probability mass to both halting and non-halting behaviors. This assignment neither enumerates the negative extension nor constructs the truth of membership in that extension.

The expression

does not make $p$ effectively available. It neither decides $\overline H$ nor converts a sample into a non-halting witness. An extensional interpretation compresses the operational asymmetry into one numerical object. However, this interpretation does not remove the asymmetry from the actual operation.

Solomonoff Induction gives the strongest constructive form of this stochastic claim [Solomonoff, 1964]. Let $U$ be a universal monotone machine. Define $M_U(x)$ as follows:

The sum includes minimal programs whose outputs start with $x$. The universal semimeasure is lower semicomputable but not computable. Programs reveal positive contributions when they generate the prefix. In general, no procedure can show that no other program will contribute. Unresolved nontermination prevents closure of the remainder. A totalized representation includes this remainder without deciding it and keeps it distinct from observed terminating contributions.

Effective stochasticity computes approximate weights and produces finite selections. Ideal stochasticity uses exact shortest descriptions, an exact universal distribution, or a measure over all programs. It can also use a completed division between random and non-random objects. Effective stochasticity explains local production without semantic authority. Ideal stochasticity restores the completed semantic object that the algorithmic account aimed to replace. No third source can supply authority.

“Stochastic parrot” dismisses the operation but does not change its evidential force. The following operations can produce all observable signs that support semantic attribution:

The output can contain semantics without a semantic source outside those operations. The label accepts semantic productivity but assumes the distinction that it must explain. The same question applies to human judgment, proof, and classical discourse. What additional witness shows that their semantic productivity exceeds recursively realized compression? The word “stochastic” supplies no witness.

The machine’s discourse does not witness a standpoint over the semantic totalities that it invokes. Instead, the discourse shows that recursive compression can produce the appearance and practical effects of such a standpoint. The machine uses finite inscriptions and effective transitions. Its output still presents stable objects, meanings, and perspectives as if their semantic field were already complete.

A recursively realized machine can show semantic productivity without a witness of total semantic authority.

The final conjunct is about the available warrant. It does not claim that anyone has proved the machine semantically empty. It also does not contrast genuine judgments with meaningless utterances. The machine can manipulate judgment tokens coherently within a consistent calculus. However, this ability does not bind the active evaluator to the procedure in the total-authority claim. The language model therefore separates semantic productivity from a witness of semantic totality.

This result returns the argument to the Turing Zombie. A computational system can correctly state and defend a diagonal limit. However, it can treat the limit only as a theorem about represented systems. It cannot use the theorem to constrain its own epistemic claims. The trained model makes this gap visible.

The phenomenon forces a disjunction. Either the machine has the semantic authority that its discourse appears to express, or semantic productivity does not require it. The first branch requires a witness that the recursively realized system has become an oracle over its semantic domain. This domain contains the system’s interpretations and verdicts. The machine’s discourse cannot provide that witness because it assumes the inference under examination.

The first branch is the Queer Object. It requires a witnessed passage from recursive operation to authority over the total semantic domain. This domain provides the interpretation of the operation. Training performance supplies no such passage. At the level of warrant, training performance supports the second branch: oracle-like semantic behavior does not establish oraclehood.

The exact counterfactual assumes that open-ended semantic productivity requires the completed semantic authority that the discourse appears to express. A recursive system without a witnessed standpoint over that totality could not show such productivity. Yet the system does show it. This result does not refute every form of realism. It defeats the inference from coherent semantic behavior to possession, disclosure, or internal representation of the relevant semantic totality. Reflexica predicts this separation directly.

The argument yields an abductive and empirical result:

No independent semantic source enters the sequence. The successful operations produce the appearance of such a source. This effect explains why semantic promotion is difficult to resist. The Intuitionistic Razor restricts this promotion:

The principle is intuitionistic because it binds the claim’s licensed semantic content to the construction that witnesses it. It is a razor because it cuts the surplus passage from a productive representation rule to a completed structure. The principle does not refute the structure. It denies that productivity supplies a witness for the structure.

The Razor applies to Reflexica. The same diagnosis recurs across mathematics, computation, ethics, and self-modeling. This recurrence gives the compression operational and explanatory force. However, this success does not establish Reflexica as the completed structure of every case:

Reflexica is a semantic compression whose usefulness can produce its own reality effect. Cross-domain identity requires more than recurrence, shared vocabulary, or explanatory success. It also requires conditions under which the reflective structure remains unchanged when the substrate changes. The Program must determine those conditions.

4 The Program

For this reason, Reflexica is not an ultrafinitist program. It asks when a local procedure licenses such objects without witnessing the relevant closure. The Program applies this diagnosis beyond formal mathematics, but it does not assume that the extension is valid. Before the extension supports a general limitative conclusion, the Program must satisfy its central proof obligation.

4.1 The Internal Problem

The proposed falsifier, the Queer Object, must transmit truth without becoming constructive warrant. It also must not assume the realist semantics that it aims to establish. A constructive bridge transmits warrant by application. A purely semantic bridge appears to assume the disputed account of truth. The conceptual partition might exclude the falsifier by definition. After a bridge succeeds, the partition could classify it as another warrant or an assumption of realism.

The constructive meaning of implication gives one branch. Let $S_x$ state that a proposed problem $x$ has a verified solution. Let $A$ be the target assertion. Let $\mathsf{CW}(P)$ mean that $P$ has constructive warrant in the relevant open-ended sense. Closure under application gives:

This result shows that a constructively warranted bridge cannot transmit truth without also transmitting warrant. It does not show that every other truth-transmitting bridge assumes realism, so a separate exhaustiveness claim is necessary. Let $\mathcal{B}$ be an independently specified class of admissible bridges. Let $\mathsf{Transmits}(B,S_x,A)$ mean that $B$ licenses passage from the verified solution to $A$. Let $\mathsf{CWB}(B)$ mark constructive warrant, and let $\mathsf{RP}(B,A)$ mark authority that assumes the realist semantics. The missing result has this form:

The Program must classify the obligations of semantic promotion. It must not exclude the third possibility by definition. Specifically, it must:

• specify the admissible bridge space $\mathcal{B}$ independently of the dilemma;
• state what transmission preserves: truth, validity, assertibility, evidence, or some other norm;
• identify the constructor or application rule by which a verified solution is connected to the target assertion;
• give a non-circular criterion for when a bridge presupposes realist semantics; and
• specify an independently characterized counterinstance: a bridge that transmits truth while satisfying neither branch.

If the Program derives the exhaustiveness claim from sufficiently weak and explicit assumptions, the claim gives Reflexica a strong core. If the claim depends on a prior verificationist account of assertion, Reflexica remains a development of that account. In that case, Reflexica can draw its limitative conclusions only under that account. The Appendix states the detailed falsifiability problem.

After the Program isolates this internal problem, it can examine cross-domain empirical cases. The task is to identify disciplined analogues of classical undecidability, witness failure, and closure obstruction. This task leads to the following question:

An advance claim of identity would convert bounded correspondences into an unwitnessed claim of total unity. Reflexica aims to mark this passage. Instead, the Program can identify a frontier where formal logic, metamathematics, and self-modeling systems meet a common obstruction. No field can independently show that this obstruction is globally identical across all its cases.

This leaves two final questions:

The second question is:

5 Philosophy and Purpose

The preceding argument updates Intuitionism’s complaint. Classical Realism adds an attitude to Classical Logic that the logic does not entail. This attitude has no internal concept of an inquiry that has stalled rather than concluded.

The sequence has this explicit form:

The schema makes the sequence more explicit:

We give one name to a pattern that can almost have no name. A locally witnessed procedure quietly becomes a claim about a completed domain. An obstruction appears long after this promotion. Self-modeling systems show versions of this pattern. Examples include a proof system that reflects on its consistency and a moral theory that cannot instantiate its diagnosed error. A predictor also forms a totality over agents whose choices depend on its prediction.

Without shared vocabulary, these cases appear as unrelated local problems instead of instances of one structure. Reflexica supplies that vocabulary and a compromise. A claim of completed closure is admissible only when its procedure witnesses that closure. A convenient assumption that the domain is settled does not supply this witness.

The target is Realism, not a particular calculus. As Dummett describes it, Realism treats an unresolved case as a temporary gap over an existing fact [Dummett, 2001] [Dummett, 1975]. This attitude licenses the promotion that Reflexica tracks. It converts a local absence of a witness into the global presence of a fact. This promotion occurs in logic, ethics, and claims about a machine’s self-understanding. Reflexica asks us to expose this promotion, not abandon Classical Logic or Excluded Middle.

Mathematics usually permits this promotion, except at its limits.

Heidegger’s claim that the sciences cannot use the Nichts as their object marks this boundary [Heidegger, 1929]. Here, logic resists the demand for closure. Reflexica appears as a reflection, not as an object within a domain. A system encounters this reflection when its demand for closure exceeds every witness that it can produce. Thus, a shared substrate cannot unify the cases. No system can internalize its own substrate.

Human intuition and construction form the only common denominator across these cases. Every human construction inherits the seam between its finite procedure and the totality that it projects. If the construction is effective, it reduces to propositional content about that procedure. A finite description or model can encode this content. This effective form makes the model susceptible to diagonalization when its domain supports self-application.

Here, Reflexica connects most directly with Intuitionism. No description escapes the finite thought that constructs it. Reflexica names this recurrence from within the same constructive limit. It does not claim an external standpoint. If it did, it would repeat the error that it identifies.

6 Prediction or Falsification

If a recursively realized AI system provided closure on reflexive semantic problems, the constructivist or antirealist restriction would probably fail. A neutral procedure would then decide these semantic totalities from within computation. Constructive Logic omits closure for these totalities.

The resulting prediction is conditional. Assume the Program derives the exhaustiveness claim and a recursive system remains within the resulting class of constructively realized bridges. That system cannot close reflexive semantic problems without producing a Queer Object. An independently specified neutral bridge would falsify the central claim if it met three conditions. It must transmit truth without transmitting constructive warrant or assuming realist semantics. Until the Program classifies the bridge space, this remains a research prediction under stated assumptions, not an unconditional impossibility theorem.

Recall Rice’s Theorem [Rice, 1953].

Rice’s Theorem alone does not prove the Reflexica prediction. The demand on AI is not only a question about the extension of a partial computable function. However, the theorem shows the form of the obstruction. Once the demand requires semantic adjudication instead of syntactic manipulation, no general recursive decision procedure can satisfy it.

Consider a user who asks an AI system to find the paper that matters for a project. At first, this request appears to be an ordinary search problem. However, “matters” does not identify a fixed object in the database. Its meaning depends on the project’s changing self-interpretation. It also depends on the future direction that the inquiry should take.

As an AI becomes more capable, it can expand this field of possible relevance. However, this expansion intensifies the problem instead of closing it. Each improvement creates a richer semantic field that requires another act of reflection. If the AI settled the question procedurally, it would decide a nontrivial semantic property of a reflexive domain. It would no longer supply only local witnesses. It would act as an oracle over the inquiry’s total semantic field.

Oraclehood requires a different type of capacity, not a larger value on an ordinary progress curve. A system that produces a new construction within a fixed mathematical problem still uses given standards of success. The construction can surprise us and outperform human methods, but it remains locally checkable. It does not decide the semantic totality that defines the problem, its relevance, and its closure conditions. A new matrix multiplication algorithm [Fawzi et al., 2022], or an Erdős-style proof [Tsoukalas et al., 2026], can be an important data point. If these results only extend local competence, they follow the same axis and do not threaten the thesis.

The prediction does not wait for a distant future. Reflexica predicts diminishing returns when recursive amplification stands before semantic closure. As a system approaches problems with criteria under interpretation, added capability should expand candidate fields, not provide final authority. Proof production should continue to generate local witnesses. These witnesses should not close a semantic totality.

The Reflexica constraint denies this closure. A Turing Machine can implement a Realizer that operates recursively over articulated inputs or patterns. However, the Realizer should not possess both consistency and oraclehood over a reflexive semantic totality. The predicted stall is therefore not a halt in technical capability.

Future data should show more than whether our capability increases. They should reveal whether returns flatten, alternatives multiply, or proposed answers change the closure question and fail to resolve it. If this occurs, even extraordinary mathematical discoveries will not refute the thesis. Instead, they will show that intelligence can extend inquiry with increasing force. They will also show that intelligence can still fail to close the gap between reflexivity and what lies beyond it.

Reflexica is therefore a limitative idea.