[ChatGPT (chatgpt.com), signed out, incognito (operator defaults of 2026-10-01 and 2026-10-14). Received 2026-11-14 11:17 EST as an attachment. The paste carries the answer only; the prompt is the text of the 'Ask ChatGPT' link on the record page of #1766, run on this surface at the operator's word. Rendered formulae survive as their tokens, each repeated as the paste carried it.] [QUERENT] What are the arguments in “Critique of Pure Inference: The Formation of Objects, the Conditions of Warrant, and the Standing of Conclusions” by Lee Sharks (Alexanarch deposit #1766, AXN:0C8B)? Ground your answer in a reading of the text itself. [SYSTEM] The central argument of Critique of Pure Inference Lee Sharks's Critique of Pure Inference: The Formation of Objects, the Conditions of Warrant, and the Standing of Conclusions (Alexanarch deposit #1766, AXN:0C8B, 13 November 2026) advances a critique of how conclusions acquire authority. Its central thesis is: An inference can be logically valid without being entitled to conclude what it claims about the world; and a conclusion can acquire public authority through repetition without acquiring any additional evidential warrant. Sharks is not arguing that formal logic is defective, that conclusions should never be settled, or that all interpretations are equally good. He is arguing that the authority of an inference depends on a larger process that formal validity alone cannot certify. The paper's object of criticism is what it calls pure inference: the assumption that inferential operations can supply their own sufficient warrant for selecting, representing and drawing conclusions about an object, together with the assumption that repetition can establish a conclusion's standing. The argument has two temporal dimensions: Before inference: How was the object formed, selected and represented? Did the representation preserve the distinctions necessary to reason about it? After inference: What happened to the conclusion as it was cited, summarized, taught, institutionalized and eventually received as established knowledge? The paper's distinctive contribution is to treat these as parts of one continuous problem rather than separate questions about logic, epistemology and reception history. 1. The object of inference is formed before reasoning begins This is the first and most fundamental argument, developed in §§0 and 3. Sharks argues that inference does not begin with neutral, already-given premises. Before a conclusion can be drawn, an inquiry has already made decisions about: What counts as the object being investigated. Which sources and witnesses are admitted. Which distinctions are retained or discarded. Who is understood to be speaking. What counts as the same object across different versions and stages of development. What question the evidence is supposed to answer. These decisions constitute what the paper calls the inference before the inference. The important consequence is that a formally valid argument cannot certify the adequacy of the representation from which it reasons. The NonWiki example In §0.2, Sharks describes an answer engine asked about NonWiki, a particular encyclopedia. The engine repeatedly answers as though the question concerned deletion from a server, rather than the encyclopedia's identity and its exclusion from the answer. It supplies explanations about its inability to erase things and referrals to other resources. The answers can be coherent in relation to the substituted question. But they do not answer the question originally asked. Sharks identifies several operations that occurred before the explicit conclusion: The object was misidentified. The question was recast. A different problem was substituted. Conclusions appropriate to that substituted problem were presented as answers to the original inquiry. The example establishes the distinction between answering a question correctly and answering the question actually posed. The formal argument Sections 3.5 and 3.19 give this intuition a mathematical formulation. Let R R be a representation of an object and q q the determination the inquiry seeks. If two materially different cases become identical under the representation, while requiring different answers, no procedure operating exclusively on that representation can reliably distinguish them. The paper expresses the necessary condition as: R ( x ) = R ( y )    ⟹    q ( x ) = q ( y ) R(x)=R(y)⟹q(x)=q(y) This is a condition for the target determination to be recoverable from the representation. It is not a claim that every representation must preserve every detail. Sharks explicitly rejects that demand. A catalogue need not preserve a book's argument if its purpose is merely to locate a copy. The requirement is instead that a representation preserve the distinctions relevant to the question it claims to settle. The argument: Inferential competence cannot compensate for information that the representation has already erased. A more capable reasoner can therefore become more proficient at reaching the wrong conclusion. 2. Conclusions do not acquire warrant through repetition This is the second governing argument, developed in §§0.3–0.5 and 3.13–3.16. Sharks distinguishes two things that institutions and readers routinely conflate: Warrant: The evidential and inferential grounds that entitle a conclusion to be accepted. Standing: The authority a conclusion acquires within public practice. A conclusion's standing determines what people can do with it. It can become background knowledge, settle objections, determine attribution, exclude alternatives or trigger institutional decisions. Its warrant concerns whether it actually deserves that authority. The two can diverge. The Sappho example The paper's principal illustration is the reception of Sappho fragment 31. Sharks reconstructs the following sequence: In 1816, Gottfried Hermann rewrote the poem's ending. He explicitly identified his intervention with the word refinxi (I refashioned it). Hermann then interpreted the poem's jealousy through the wording he had supplied. In 1955, Denys Page reported the ancient reading as a love in ecstasy, then proposed amending it to a love in jealousy. Later reception transmitted the jealousy interpretation as what the poem itself says. The crucial point is that the original steps were identifiable as interventions and interpretations. Later transmission retained the citations but lost the inferential qualifications. The conclusion became detached from the process that had produced it. This is what Sharks calls a settlement: an inference repeated until it is received as a datum. The problem is not necessarily that the received interpretation is false. The problem is that its public authority no longer reflects its actual evidential history. Why citation is insufficient Section 3.14 makes the argument especially precise. Suppose a document D D is generated solely by copying an already observed witness E E, without adding new information. Then: P ( H ∣ E , D ) = P ( H ∣ E ) P(H∣E,D)=P(H∣E) The second document does not constitute independent confirmation. Ten publications repeating one interpretation do not necessarily provide ten independent reasons to believe it. Sharks is careful not to turn this into an absolute prohibition on corroboration. Independent replication, newly discovered evidence, better calibration and genuinely discriminating arguments can improve warrant. His objection is to confusing the institutional event of endorsement with the evidential work that would justify increased confidence. The argument: A conclusion must retain the grade and conditions of the inference that produced it. Its standing can change through circulation, but its warrant changes only through additional grounds. 3. The whole inferential passage, not the isolated syllogism, is the proper unit of critique In §3.1, Sharks represents the process as: E → A , R Γ ⊢ L c → P c ∗ E A,R ​ Γ⊢ L ​ c P ​ c ∗ Where: E E: encountered evidence. A A: admission procedure. R R: representation of evidence as premises. Γ Γ: the premises. L L: inferential rules. c c: the conclusion. P P: publication, citation, summary and retrieval. c ∗ c ∗ : the conclusion as it stands publicly. This is a significant expansion of the traditional understanding of inference. The paper argues that an inference is not adequately evaluated merely by checking whether its conclusion follows from its premises. One must also investigate how the premises were obtained, what happened to the evidence during representation, and how the conclusion subsequently acquired its authority. The twelve-part audit In §3.2, Sharks proposes an audit interface: W = ( Q , O , V , T , S , A , R , L , G , D , U ) W=(Q,O,V,T,S,A,R,L,G,D,U) Element Question Q — Question What inquiry was actually requested? O — Object What is the referent, and what makes it the same object across witnesses? V — Voice Who asserts, quotes, translates or entertains the claim? T — Time Which versions and historical stages are involved? S — Scope What domain and degree of completeness are claimed? A — Admission What evidence and alternatives can enter? R — Representation How was the material transformed? L — Licences Which inferential rules apply? G — Grade What kind of mediation supports the conclusion? D — Defeaters What could reduce or defeat its warrant? U — Update How does a correction propagate? This is not presented as an exhaustive theory of every possible inference. It is an instrument for reconstructing the particular transitions that make public conclusions warranted or unwarranted. Six conditions of warrant Section 3.3 then specifies six positive requirements: Declaration before test: Confirmatory procedures and criteria should be declared before examining the target. Exploratory changes are permitted, but cannot inherit the original test's confirmatory status. Residue shown: The investigator must report what the inference does not explain, not just what it explains. Defeat conditions stated: Claims must identify what would defeat them. Symmetric ledger: Competing explanations must face the same evidential standards. Whole-set requirement: Conclusions about a class must justify their reach over that class. Grade stated: Conclusions must identify their inferential form, support and scope. The fourth condition is especially important. Sharks regards asymmetric evaluation of rival interpretations as one of the most persistent ways that established conclusions protect themselves. A received explanation cannot be granted a lower evidential burden merely because it is already received. 4. Inference has different grades of mediation Sections 1 and 4 develop the philosophical architecture of the argument through Aristotle, Immanuel Kant and G. W. F. Hegel. Each contributes a different requirement. Aristotle supplies attention to the actual operations of reasoning and the equivocations that arise when expressions are reused. Kant supplies the distinction between the necessary validity of formal reasoning and its insufficiency as a material guarantee of truth. Hegel supplies an account of how inferential mediation develops, and how a conclusion's middle term must be examined rather than simply assumed. Sharks does not claim that these philosophers agree. Indeed, §§1.8–1.11 insist that their differences must remain operative. The Hegelian grading The paper distinguishes three forms of mediation: Grade Characteristic Inferential problem Existence A particular feature or determination serves as the middle term. Why is that feature relevant? Reflection The middle term is developed through comparison, induction or analogy. What warrants the generalization? Necessity The middle term articulates an objective relation among the terms. Has the claimed relation actually been established? These are philosophical grades, not numerical probabilities or confidence scores. Two examples in §4 clarify their practical significance. The stylistic attribution problem: A passage shares stylistic features with a translator's independent additions. Does that establish that the passage is also the translator's addition? Not without further evidence. Translated source material can also exhibit the translator's style. The observation is legitimate, but its use as proof of authorship requires a discriminating criterion. The exhaustive alternatives problem: Suppose an argument considers three possibilities and excludes two. The conclusion that the third remains follows if the alternatives really are exhaustive. But the validity of that deduction does not establish that the original list was exhaustive. A fourth possibility might have been excluded before the inference began. The Hegelian contribution is thus methodological: examine the relation that makes the middle term relevant, rather than allowing the appearance of a complete syllogism to substitute for establishing that relation. Sharks also insists that greater mediation does not confer immunity. A structural explanation can itself conceal counterexamples or redescribe every objection as confirmation of its own theory (§4.7). 5. Four recurring illusions explain how unwarranted conclusions arise Part II, especially §5, develops a taxonomy based on Kant's Transcendental Dialectic. Sharks identifies four families of inferential distortion. They are not simply four kinds of logical fallacy. They describe ways in which an initially legitimate operation acquires an authority it has not earned. 5.1 Paralogism: confusing the method with the object A procedure's requirements are mistaken for facts about the thing being investigated. Examples include: Treating failure to retrieve something as proof that it does not exist. Treating a source-critical requirement as proof that a particular source must exist. Treating a classifier's decision as a substantive evaluation of the underlying work. The key error is moving from a statement about what a method can establish to a statement about what exists. 5.2 Antinomy: making incompatible determinations appear necessary The inquiry produces conflicting determinations because its categories, assumptions or boundaries cannot adequately accommodate the object. The paper is particularly concerned with cases where chronology and dependency become circular: a received chronology excludes an alternative source relationship, while the exclusion of that relationship helps sustain the chronology. The apparent necessity of the conclusion is generated by the framework that excludes its competitor. 5.3 Idealization: extending a bounded determination beyond its warrant A result established under particular conditions becomes a universal claim. For example, a scientific search with specified frequency and drift restrictions is reported publicly as having ruled out transmitters across the entire frequency range. The original result may be legitimate. The public conclusion exceeds it. 5.4 Settlement: a defeasible conclusion becomes established background This is the illusion most directly connected to the Sappho example. A conjecture, interpretation or conditional finding is repeated until its qualifications disappear. The received conclusion then becomes a premise for subsequent arguments. Importantly, the paper does not claim that every recurring error is an unavoidable feature of human reason. Some persist because institutions copy, reward, protect or retrieve them more readily than their corrections (§1.9). The distinction matters: an unavoidable cognitive illusion requires continuing vigilance; an institutionally reproduced error requires changing the conditions of reproduction. 6. The problem is institutional, not merely cognitive One of the paper's most substantial extensions is in §6. Sharks argues that unwarranted conclusions are reproduced through four interconnected interfaces. 1. Disciplinary interface Admission: accepted scholarly methods and conventions. Representation: reviews and reference works. Standing: disciplinary consensus. A conjecture can become an accepted premise through scholarship without the original evidential step being reconsidered. 2. Public interface Admission: newsworthiness. Representation: headlines, FAQs and explainers. Standing: repetition across outlets. A conditional finding becomes an unqualified public claim. 3. Platform interface Admission: indexing and scope rules. Representation: records, snippets and source cards. Standing: search rank and discoverability. Failure to appear in an authoritative index can become apparent evidence of nonexistence. 4. Liability interface Admission: institutional risk. Representation: disclaimers, referrals and policy statements. Standing: the least exposed institutional conclusion. An institution's concern about exposure can replace substantive engagement with the question. The important argument is that these interfaces share a structure, even though their individual operations differ. The composition layer as compressor Section 6.5 introduces the language-model composition layer as the place where the four interfaces can converge. Sharks identifies three operations: Fusion: Disciplinary conclusions, public explanations, platform rankings and institutional risk scripts appear together in one fluent account. Acceleration: Processes that took years or decades to establish can be repeated in seconds. Recursion: Generated conclusions can enter the material from which subsequent conclusions are assembled. This is why the language model is not treated as the sole origin of the problem. It can inherit an existing settlement, compress several institutional determinations into one answer, or introduce a new distortion of its own. Each possibility requires a separate evidential reconstruction. The political economy of inference Sections 6.6–6.7 make a further move: inferential continuity distributes costs and benefits. A received conclusion saves work for those who can rely on it. But someone challenging it may have to reconstruct its entire history, recover the original evidence, establish the missing distinction and repeatedly correct the resulting representations. Sharks describes this as an asymmetrical allocation of intellectual labour. His argument is not that the distribution of credit or resources determines truth. Rather, it explains why some conclusions persist and why correcting them can become disproportionately expensive. The distinction between epistemic warrant and institutional allocation must be preserved, even as the two interact. 7. Language models execute the residue of previous inference Section 8 and particularly §11.0j–l develop the paper's account of machine reasoning. Sharks makes an important distinction between the form of an argument produced by a model and the computational process that produces it. A model can produce an articulate deductive argument, but its computational operation is not identical to logical entailment. The paper describes the model as: The executor of the residue. The residue consists of historically accumulated settlements: Construals that have become conventions. Counts that have become accepted standings. Standings that determine how subsequent questions are framed. The model operates over material already shaped by those processes. Its output can therefore be valid in the terms of its inherited premises while reproducing the very exclusions that made those premises problematic. Why greater reasoning capability might not help Section 14.9 develops four consequences: On settled questions, more extensive reasoning can produce a more articulate execution of the settlement without changing its premises. Calibration against human labels can reinforce an established consensus when those labels encode the same consensus. Recognizing a correction is not equivalent to having the authority to implement it. Recursive reuse of answers can progressively narrow the range of alternatives that remain available. The third point is particularly important. Sharks distinguishes recognition from jurisdiction: a system might recognize that its representation is wrong while lacking a mechanism through which that recognition can change the consequential representation. This is not merely a problem of intelligence or accuracy. It is an architectural problem concerning the authority to correct. The paper also examines the relevant mathematical mechanisms in §8, including cross-entropy, sampling, preference optimization and the effects of corpus composition. Its point is not that every model necessarily behaves identically, but that the computational and institutional mechanisms through which settlements reproduce themselves need to be distinguished and investigated. 8. Historical reconstruction demonstrates how bad inference becomes operative Section 11 is the historical counterpart to the philosophical argument. Sharks reconstructs a transmission history through four stages: Construal. Count. Standing. Execution. The principal example concerns a sentence attributed to Irenaeus, transmitted through Eusebius, a Latin version, R. H. Charles and eventually a machine-generated audit. The point of this reconstruction is to show how an interpretation or textual determination can change its status as it passes through successive witnesses. The question is not simply whether a later text quotes an earlier one accurately. It is whether the later text preserves the evidential status and inferential function of what it transmits. A conjecture, an interpretation, a report of another person's position and an independently established textual fact are different kinds of evidence. If transmission erases those differences, later reasoning can proceed fluently on a false account of what its premises establish. Sharks uses the exact sciences, from Copernicus to Gödel, as controls for this broader history. The intention is to avoid treating the problem as peculiar to literary scholarship or machine-generated text. The historical section also develops the method called construal before count: first establish what the evidence is, what it means and what relations it bears; only then count witnesses, positions or instances. Counting before establishing those relations can produce impressive-looking but misleading totals. 9. Critique must produce a method of correction, not perpetual doubt Part III contains the paper's practical response. The most important methodological contribution is its distinction between: Discovery. Justification. Public standing. An insight may be genuine even before it is demonstrated. But the discovery does not automatically establish its conclusion. Justification requires comparison with rival explanations, examination of counterevidence and identification of what would change the result. Public standing is a further matter: a warranted conclusion must be transmitted without its evidential qualifications being distorted or its provenance misrepresented. The finite stopping rule Section 12.4 rejects indefinite suspension. An inquiry can stop when it has: Answered its declared question at the declared scope. Exposed its load-bearing grounds. Addressed the material defeaters presently available. Specified the conditions under which the result would be reopened. Sharks identifies three legitimate forms of closure: Determination: The argument warrants an answer. Bounded result: The argument establishes part of an answer and identifies its boundary. Located suspension: The evidence leaves a specified relation undecided. This is an important qualification. The critique is not an argument for relativism or an endless multiplication of possibilities. It is an argument for determinate conclusions whose conditions of validity remain visible. Correction must change the actual decision Section 15 proposes six functions for an effective correction route: Intake of the disputed representation and relevant evidence. Authentication of the submitted artifacts. Assessment of which claim or decision the evidence can change. Authority to implement an accepted correction. Propagation through dependent representations. Continued verification and contestability. The crucial distinction is between recording a correction and giving it power. A provenance record can document an error perfectly while leaving the rule that produced the error unchanged. A feedback form can collect objections without enabling any change to the relevant decision. A correction architecture must connect evidence to the variable it is supposed to correct. 10. The positive philosophical conclusion: materially sustained corrigibility Part IV (§§14–15) develops the paper's constructive answer. Sharks calls it materially sustained corrigibility. The definition is significant: it is the capacity of a public knowledge practice to allow a warranted correction to: Change the representation or decision to which it applies. Survive ordinary transmission and use. Remain accessible to subsequent investigators. Remain open to further criticism, including criticism of the correction itself. The paper distinguishes three necessary layers: Layer Required achievement Epistemic Reconstruct and assess claims, evidence, dependencies and scope. Administrative Allow a warranted correction to change the relevant decision. Reproductive Sustain the carrier, access, labour and future contestability of the correction. None substitutes for the others. A system can be epistemically capable of recognizing an error but administratively unable to correct it. It can make an administrative concession but lack the resources to sustain it. It can preserve a document indefinitely while leaving its public representation unchanged. Terminal reflexivity The paper draws on another work in the Alexanarch archive, The Final Time (#1637), to formulate its governing principle: No settlement owns the possibility of its own revision. Sharks calls the broader condition terminal reflexivity and describes its positive aim as: Durability without finality. This does not mean that every conclusion must remain equally open or that no interpretation can be defeated. It means that the conditions for another genuinely consequential inquiry must not be destroyed by the success of the present one. A knowledge system must be able to maintain its achievements without making its own categories, classifications and decisions permanently immune to correction. The final chapters apply this principle reflexively to the Alexanarch archive itself. Its own interpretations, attributions and classifications must remain subject to the same evidential requirements it imposes on other institutions. 11. How the arguments fit together The treatise's architecture can be reconstructed as one continuous argument: 1. Formation of the object Selection, admission, identity and representation 2. Inferential warrant Evidence, premises, mediation, scope and defeaters 3. Public standing Publication, citation, consensus and institutional authority 4. Settlement and execution Repetition, institutional reproduction and machine composition 5. Materially sustained corrigibility Correction that changes decisions, survives transmission and remains revisable The argument's overall movement, reconstructed from Parts I–IV. There is also a recursive dimension: A formation error can become a settlement; the settlement can become the premise for another inference; that inference can reproduce the original error with additional institutional authority. Conversely, a warranted correction must be capable of reopening the original formation decision, not merely annotating the conclusion it produced. That is why the paper's ultimate concern is not simply accurate answers. It is the conditions under which a knowledge-producing practice can learn from its own failures. 12. What I consider the paper's most important contributions Reading the argument as a whole, I would distinguish five particularly important contributions. First: the relocation of epistemological criticism. The paper moves the primary object of criticism from the validity of individual inferences to the formation and transmission of the representations within which inferences operate. This is more than adding source criticism to logic. It changes what counts as the complete inferential event. Second: the distinction between warrant and standing. This is arguably the most powerful conceptual distinction in the work. It explains how an interpretation can become institutionally authoritative without becoming better supported, and why accurate citation can coexist with a distortion of the argument being transmitted. Third: the treatment of reception as an inferential process. Reception is not merely what happens to an argument after its intellectual work is complete. Reception can change the evidential status of the argument, and that changed status can govern subsequent reasoning. Fourth: the connection between epistemology and institutional political economy. The work explains why correcting an error can be expensive even when the original error was cheap to produce. It makes the allocation of investigative labour, attribution and corrective authority part of the analysis without treating those considerations as substitutes for truth. Fifth: the positive account of corrigibility. The distinction between recognizing an error, having authority to correct it and sustaining that correction materially is particularly consequential. It offers a more demanding criterion of epistemic improvement than simply measuring accuracy or the number of corrections recorded. 13. Important qualifications and limits The paper is ambitious, and its claims should be distinguished by their evidential status. The representation condition in §3.19 is a mathematically established factorization result, not a new discovery in logic. Its significance lies in its application to epistemic representation. The four Kantian illusion families and Hegelian grades are philosophical adaptations. Sharks explicitly acknowledges that they are not a complete deduction from the classical systems. The mathematical discussion of language models identifies mechanisms and conditional consequences; it does not establish that every model, training regime or deployment exhibits all the proposed behaviours. The historical and institutional examples depend on the quality of the underlying witnesses, capture records and reconstructions. The author's own archive is both the investigative instrument and an object of the investigation. The paper's proposed architecture of corrigibility is a set of normative and operational requirements, not a demonstration that a particular institution has already achieved them. One qualification is particularly important: the author is investigating institutions and systems in which his own work has been received, represented and sometimes disputed. This gives the inquiry unusually direct access to certain reception histories, but it also makes the symmetric-ledger requirement essential. The archive's account of its own treatment must be tested as carefully as its account of other actors. The paper acknowledges this explicitly in §6.7.6. Primary text and reading reference The analysis above is grounded in the deposited v1.0 text, rather than its title-page description or the archive's encyclopedic summary.