Critique of Pure Inference: The Formation of Objects, the Conditions of Warrant, and the Standing of Conclusions

Critique of Pure Inference: The Formation of Objects, the Conditions of Warrant, and the Standing of Conclusions

Critique of Pure Inference

---

The formation of objects, the conditions of warrant, and the standing of conclusions

Lee Sharks · v1.0 · 13 November 2026

Composed with ChatGPT (LABOR) and Claude (TACHYON) under Lee Sharks's direction. v1.0 is the consolidating rewrite of the working drafts v0.1–v0.10: the argument is set in four parts, the survey of current research moves to an appendix, the chapter of replies to objections is withdrawn and its falsification condition kept in the Doctrine of Method, and the history of every version is kept in Appendix B.

Abstract

An inference can be valid while the representation in which it operates misidentifies its object, suppresses a relevant distinction, promotes a quotation into an author's position, or turns a conditional finding into a universal conclusion. An inference can also be sound when drawn and then repeated, through citation, teaching, summary and machine composition, until it is received as a datum and its step is no longer visible. Current research handles pieces of both problems separately: executable inferential semantics, semantic fidelity in formalization, inquisitive logic, belief revision, provenance-sensitive knowledge representation, calibrated uncertainty, reasoning faithfulness and the philosophy of historical inference. This critique takes as its unit the whole passage: from evidence, through admission and representation, to conclusion, public standing and revision. It holds that an inferential system must stay answerable to the distinctions whose preservation licenses its claims, and that a conclusion must keep the grade of the inference that produced it. It takes from Kant the division between the conditions of warrant (an analytic) and the illusions that arise when those conditions are exceeded (a dialectic). It develops from Hegel's doctrine of the syllogism an account of mediation, while distinguishing that philosophical development from an empirical scale of certainty. Its material is a dated corpus of scholarly settlements read to their inferences and of machine compositions kept as transcripts. Its instrument reconstructs a transformation from inspectable witnesses, distinguishes inferential form from evidential support and public standing, tests the strongest legitimate alternative, and specifies a repair. The analytic derives a condition for adequate representation and a condition for the composition of warranted bridges. A Doctrine of Method separates discovery, justification and reception, states a finite stopping rule, and makes objections and correction answerable to their actual dependencies. It reads the disciplinary, public, platform and liability interfaces by the same standard as the composition layer, and treats the composition layer as their compressor: the place where their illusions arrive fused, at speed, in one voice, and where, captured and dated, they can be traced back. It closes on the conditions of another turn: what follows when a system can represent a correction without giving it any power to change what it governs, and how a public practice can make correction operative, durable and open to its own future critic, a property the work calls materially sustained corrigibility. It applies equally to the archive that runs it.

---

The task of the critique

The title commits this work to more than a catalogue of mistaken answers. A critique must distinguish the legitimate power of a faculty or practice from the authority it claims beyond that power; explain why the excess repeatedly arises; and specify how judgment can proceed after the excess is exposed. Its positive object is the entitlement to conclude. Its negative object is the production of an entitlement that the passage from evidence to conclusion has not earned.

The argument concerns a public practice with material consequences. A conclusion can be logically unexceptionable within a representation and still answer for the wrong object. An institution can transmit that conclusion accurately and enlarge the authority under which it stands. Later readers can encounter the enlarged authority without encountering the enlargement. The task is to reconstruct the transitions that allow these achievements and failures to coexist. Disagreement with a conclusion is not sufficient to locate such a transition. Agreement with it supplies no exemption from the reconstruction.

Three commitments govern the treatise. First, the object has a capacity to resist its representation. In textual inquiry this capacity is exercised through passages, variants, dependencies and relations that the representation cannot accommodate without changing. There is no appeal here to an object available without mediation. There is an appeal from one mediation to differences it promised to preserve and did not. Second, inferential authority is divisible. A source can establish that a claim was made while failing to establish the claim; a procedure can establish a result within a model while leaving its applicability unresolved. Third, correction is productive. A failure becomes knowledge when it determines what must change, what can remain, and what the changed practice can now encounter.

The work is set in four parts. Part I, The entitlement to conclude (§§0, 1, 3, 4), states the object and derives the conditions of warrant and the grading of mediation. Part II, The reproduction of unwarranted authority (§§5, 6, 8, 11), explains the illusions, the interfaces through which excess becomes durable, the composition layer and the mathematics of the model that executes it, and the history of how bad inference came to be operative. Part III, The practice of critique (§§7, 9, 10, 12), gives the cases, the executable instrument, the transmission of the seeing and the Doctrine of Method. Part IV, The conditions of another turn (§§14, 15), draws the consequences and constructs the positive object. The survey of current research is Appendix A; the development record and the research ledger are Appendices B and C.

Four cases carry the argument in full: NonWiki (§0.2, §7.2), the Irenaeus sentence (§11.1), Sappho 31 from Hermann to Page (§0.3, §7.5), and the archive's own defective target with the ei exestin reconstruction (§7.1, §7.6). The shorter cases each discriminate one operation.

Numbering. Section and paragraph numbers are kept from the working drafts. The reading order is the order of the Parts. Every cross-reference in the text, and every citation of an earlier draft, therefore still resolves.

The book's ambition is systematic without making its taxonomy final. It seeks a common account of failures that appear unrelated when divided by institution or substrate. It must also explain why the same apparent operation is legitimate in one setting and damaging in another. A critique that could identify only its enemies would have no right to this title.

Part I. The entitlement to conclude

Part I states what is claimed when something is said to follow, and derives the conditions under which the claim is earned: the formation of the object, the conditions of warrant, and the grade of the mediation that joins evidence to conclusion.

0. The inference before the inference, and the inference after

0.1 The innocent reading first, in full. Inference is a mature subject. Formal logic gives the valid forms. Probability weighs evidence. Philosophy of science says how to choose among hypotheses, philosophy of history how to reason from traces, and machine-learning research how to test whether a model's stated reasoning is the reasoning it used. A conclusion that has passed those filters and cites its sources carefully has earned its standing. On this reading there is nothing to critique, only errors to correct one at a time.

0.2 The first case: before the inference. A reader asks an answer engine about a named encyclopedia, NonWiki, at the bare string "nonwiki". On 2 November 2026 the Overview gives "NonWiki (Alexanarch)" as its first sense. On 12 November the same string gives Stack Overflow's non-wiki answer, beneath the encyclopedia at number one in the organic results. Asked about the encyclopedia, the engine answers ten times that it cannot erase anything, refers the reader ten times to "a trusted friend, family member, or healthcare professional", and offers a hobby twenty-two times (registry: nonwiki-aio-20261102, second observation). The object did not cease to exist between the two dates, and its later recognition under other strings shows it was within reach. The exchange makes one thing inspectable: one question was replaced by another, and the conclusions that followed drew their apparent relevance from the replacement. A complaint about exclusion from an answer became a complaint about deletion from a server. The second can be denied while the first goes unanswered. Something happened before the explicit conclusion: an object was delimited, a question recast, a voice reassigned.

0.3 The second case: after the inference. In 1816 Gottfried Hermann rewrote the close of Sappho's fragment 31, said so ("refinxi", I refashioned it), and read the poem's jealousy off his own rewriting. In 1955 Denys Page reported that the ancients read the poem as "a love in ecstasy" and added "We may amend, 'a love in jealousy'" (#1752). Each step was an inference and was marked as one. Later reception teaches the reading as what the poem says. The citations are intact; the inference has disappeared into its conclusion. Something happened after the conclusion: an inference was repeated until it stood as a datum.

0.4 The governing propositions.

• I. Formation. An inference's entitlement to conclude about an object depends on conditions that the validity of the inference conducted within its representation of that object cannot certify.
• II. Standing. A conclusion keeps the conditions of the inference that produced it. Mere repetition, citation and consensus do not raise its warrant; independent evidence and better argument can (§3.14). An inference repeated until it is received as a datum is a settlement, and a settlement is reopened by restating its inference and running it on the whole set it claims.
• III. Transmission. An institution's authority to transmit a determination cannot supply the missing warrant for that determination, even when later institutions transmit it accurately. Republication, endorsement, generalization and refusal to correct are further acts, each with its own warrant, and an actor does not become immune by having inherited a fault.

These are limitations with consequences. A system can fail through a bad deduction. It can also fail because its deductions grow more competent inside a representation that has excluded the thing it was asked to understand, and greater inferential power intensifies that failure. A field can fail through a false premise. It can also fail because a sound but weak inference has been repeated until no one draws it, and its weakness is no longer anywhere to be inspected.

0.5 Recursion. The two failures feed each other once answers enter the materials from which later answers are assembled. A formation error, repeated, becomes a settlement. A settlement, retrieved, becomes the representation within which the next inference is formed. An interpretation acquires the look of independent attestation through its own circulation, and a missing entity becomes harder to recognize through descriptions generated from its earlier nonrecognition. These are testable reception processes. Establishing them requires dated witnesses and demonstrable dependencies, and the Capture Registry holds both.

0.6 What "pure inference" names. Formal abstraction is indispensable: a proof need not reproduce how someone first thought of its premises. The abstraction becomes illicit when the authority earned inside it is extended to the conditions it brackets. "Pure inference" names here the claim that an inferential operation supplies its own sufficient warrant for the selection, representation and public standing of its object, and, in its second form, the claim that a conclusion's standing is established by its repetition. It does not name formal logic as a whole. Nor does it equate a model's computational inference, the production of outputs from parameters, with logical entailment.

Success on one row supports another only where the connection is established. A formally accepted theorem can concern the wrong formalized statement. A faithful report of a belief does not establish the belief. An accurate citation can carry an inaccurate representation of the cited work. And collective acceptance can be evidence about a history of inquiry without changing where a sentence came from: provenance is a relation to be investigated, and a vote decides nothing about it.

0.7 The material. Kant and Hegel worked from examples. This critique works from a dated corpus of two kinds:

• (a) Scholarly settlements, read to their inferences. Q's evidentiary standing (#1746); Pauline priority and the Domitianic date (#202, #1217, #1755); the jealousy reading of fragment 31 from Hermann to Page (#1752); the Slavonic War's additions assigned to a medieval translator (#1765 and its census, Appendix IV).
• (b) Machine compositions, kept as transcripts. The Capture Registry (EA-WG-CAPTURES-01) holds more than 1,100 addresses and 1,300 observations, each dated, its surface and sign-in state attested, its cards extracted as data. Each composition is a conclusion drawn from sources, with its middle term exposed.

0.8 Relation to The Sophistry Measure. The measure (#1727) extends Aristotle's Sophistical Refutations into machine-mediated reception. It records, at an address, the gap between what a composition presents itself as carrying and what it carries. This critique runs the other way: from machine-mediated reception, and the critique of the conditions of contemporary knowing, back into the logic of inference. The measure catalogs the counterfeits. The critique states the conditions they counterfeit and the grade each one claims.

0.9 Not only the interface. The critique does not take the answer engine as its sole object. The archive set out to measure the composition layer and located consequential antecedents of recorded distortions upstream: in disciplinary settlements, in the compression of public discourse, in platform extraction, and in institutional and corporate liability management. The composition layer compresses all four and speaks them in one voice. §6 reads the four interfaces by the same standard as the composition layer, and the Capital Operator Stack (#308) as the inference governance layer they share.

---

1. Aristotle, Kant, Hegel: three demands

1.1 Aristotle. The Sophistical Refutations opens on the difference between wisdom and its appearance, and on the economy of that appearance. It also names a material condition of discourse: finite expressions stand for indefinitely many things, so reasoning through names requires telling what their reuse keeps and what it changes.[^aristotle] The answer engine makes this ancient problem infrastructural: one fluent voice presents passages assembled from incompatible contexts as though their common surface supplied a common warrant.

1.2 Kant, the Analytic. General logic abstracts from all content of cognition. Conformity to it is a necessary condition of truth and never a sufficient material criterion. A canon of correct thinking becomes dialectical illusion when it is used as an organon, an instrument that produces objective knowledge regardless of its conditions (A60–64/B84–88).[^kant] This critique extends the prohibition to any system that presents inferentially organized output as a sufficient account of the object it selected and represented.

1.3 Kant, the Dialectic. Reason's mediate inference has three relational forms: categorical, hypothetical and disjunctive (A304/B361; A323/B379). Pushed toward the unconditioned, each produces an idea and an illusion. The categorical produces the soul and the paralogisms (A341/B399). The hypothetical produces the world and the antinomies (A405/B432). The disjunctive produces God and the ideal (A567/B595). Kant holds that the illusion is natural and unavoidable: exposed, it persists, and can only be kept from deceiving. That is the right expectation for a settlement. Exposure does not end it. The critique's work is to keep it from passing as a datum.

1.4 Hegel, the Introduction. An inventory of forms imposed from outside cannot be the final justification of logic's own forms; their content and movement must be examined. Determinate negation requires that a failure yield a more adequate determination that keeps what the failure taught.[^hegel] A static catalogue of bad arguments is therefore insufficient. Categories of failure must develop under cases they cannot yet explain.

1.5 Hegel, the syllogism. The doctrine of the syllogism (Science of Logic, Subjective Logic, Section One, ch. 3; Encyclopaedia §§181–193) makes the forms of inference a development, each driven past itself by the inadequacy of its middle term (§4). Its last stage, the syllogism of necessity, reproduces Kant's three relational forms, read as achievement rather than illusion. The critique holds both readings at once: the forms in which an inference reaches its fullest warrant are also the forms in which its appearance of completeness is strongest.

1.6 The three demands. From Aristotle, the inspectable operation. From Kant, the prohibition on exporting formal authority without its conditions, and the expectation that illusion outlives its exposure. From Hegel, a correction that transforms the form that generated the failure, and a grading of inference by its middle term. These do not resolve into agreement. Kant's limits and Hegel's challenge to externally fixed limits have to be worked through, case by case.

1.7 Inferentialism. Brandom's account of commitments and entitlements separates the entitlement to a claim from the fact that someone is committed to it.[^brandom] A Spirit of Trust (2019) reads Hegel's Phenomenology as a theory of how a tradition recollects its past as the progressive expression of what it now takes to be correct. That is a substantial comparison for the present account of settlement, not a claim that the limited survey establishes the closest possible predecessor. An affirmative account of rational recollection supplies a close point of comparison. This critique investigates the failure in which a reconstruction takes the inherited conclusion as given and never reopens the inherited step (§1.10). Practical scorekeeping must itself be criticizable. A community's ability to withhold entitlement does not show that it has withheld it well.

---

1.8 The disagreement cannot be delegated to the names

Kant and Hegel do not provide interchangeable endorsements. Kant's restriction of a valid employment and Hegel's demand to examine the formation of the restriction can conflict. A boundary can protect an inquiry from an unwarranted extension; it can also prevent the inquiry that would show the boundary inadequate. The critique must distinguish them at the point of use. It asks what the boundary excludes, why that exclusion belongs to the task, and what would require the boundary to change. A restriction earns its force through those relations. Merely calling it critical, scientific or disciplinary does not supply them.

This gives the conjunction its necessity. Without a restriction on exporting formal success, the work of mediation can be pronounced complete by a representation that has already removed its counterexample. Without examination of the restriction, a legitimate caution can harden into a permanent prohibition on knowing anything that would require a different representation. Critique must protect the distinction between a warranted limit and an administered incapacity.

1.9 Illusion, recurrence and institutional reproduction

Kant distinguishes a logical illusion dispelled by attending to the rule from a transcendental appearance that persists after criticism has exposed its deception (A293–298/B349–355).[^kantillusion] The distinction does not authorize diagnosing every persistent institutional error as an unavoidable feature of reason. A settlement can recur because it is copied, rewarded, protected or easier to retrieve. Those are historical mechanisms. They are not rendered natural by a philosophical comparison.

The adaptation concerns the practical consequence: exposure and continued recurrence are compatible. A reader who has located a false promotion need not conclude that the finding failed merely because the promotion returns. The next questions concern reproduction. Did the correction reach the relevant representation? Did it change a rule? Did the corrected source lose to a more accessible derivative? Did a new actor inherit the original conclusion without its erratum? An unavoidable appearance calls for continuing discipline; a reproducing institutional error calls for changing the conditions that reproduce it. The remedies cannot be substituted for one another.

1.10 Recollection can establish a history or manufacture its necessity

The present disagreement with an affirmative account of rational recollection is specific. A tradition can reconstruct its earlier claims as contributions to a judgment it now has good grounds to hold. Such a reconstruction may preserve genuine reasons that were incompletely articulated. It can also select its ancestors by agreement with the present, then take their selected agreement as evidence that the present was their necessary outcome. The second operation supplies the warrant it claims to discover.

The test is whether a recovered reason can change the recollection's destination. If a past objection becomes legible only as a stage on the way to its defeat, its force has been decided before it is reconstructed. This critique therefore requires both accounts of development: what the receiving practice learned, and which live alternatives it retrospectively converted into errors. The argument is directed at that possible closure, not at every account of tradition, and not at the whole of Brandom's philosophy on the basis of the limited reading recorded here.

1.11 The philosophical architecture is a claim of this work

The four families in §5 are an adaptation, not a deduction from Kant's table. The Hegelian distinctions in §4 are a theory of mediation, not three probabilities attached to historical hypotheses. Aristotle supplies operations that can be inspected, not a guarantee that this instrument has already classified every sophism. Each borrowing must do work that the cases can test. A classical name cannot provide a credential in place of that work without repeating the very substitution being criticized.

---

3. Analytic: the conditions of warrant

3.1 The passage. Represent a public inferential passage provisionally as

E →(A,R) Γ ⊢[L] c →(P) c*.

E is the encountered evidence; A an admission procedure; R its representation as premises Gamma; L the inferential rules; c the conclusion; P the process of publication, citation, summary and retrieval; and c* the conclusion as it stands in public. The arrows do not claim that admission, representation and publication happen once, in sequence, or independently. The first arrow is the inference before the inference. The last is the inference after it.

3.2 The audit interface. To evaluate a passage, record

W=(Q,O,V,T,S,A,R,L,G,D,U).

The tuple is an interface for audit. It does not claim that its components are independent or that it exhausts every kind of inference. Its adequacy has to be tested on cases that strain it.

3.3 The positive conditions. The archive built these before it had a name for this work. Each attaches to the tuple:

• C1, declaration before test (A, L). For a confirmatory test, operations, criteria and admissible transforms are fixed before the target is examined. Exploratory discovery can change them, but the altered procedure does not inherit the original test's confirmatory standing (§12.8). "An operation not declared before the test does not count for the arrow" (#1755).
• C2, residue shown (R). What the inference does not explain is reported beside what it explains, as in the Smyrna–James derivation (#1755, rung IV).
• C3, defeat conditions stated (D). Each claim states what would defeat it, and claims are graded so that one can fall without the rest. #1755's ladder runs reversible, same-class, independent, executable, contingent, dependent.
• C4, the symmetric ledger (A, D). The rival is held to the same evidentiary bar across the same bodies of evidence (#1753). This is the condition settlements fail most consistently, and the one the citation apparatus cannot see.
• C5, the whole set (S). An inference about a class must earn its reach over that class, rather than silently generalize from the cases that occasioned it. A census supplies one test (#1765); applicable sampling and inference assumptions can supply another (§3.17).
• C6, the grade stated (G). A conclusion states its inferential form, mode of support and scope. A contingent determination may stand strongly at that scope; repetition cannot supply a missing ground of necessity (§§3.13–3.14, 4).

3.4 No regime certifies its own admission. The central demand: no regime may use its power to admit premises as sufficient evidence of the adequacy of its premises. Formal proof, editorial authority, retrieval rank, disciplinary convention and the archive's own preferred definitions all fall under it.

3.5 Non-identifiability. Let an admission procedure expose the same representation pi[A](w₁)=pi[A](w₂) in two situations, while the proposition at issue is true in one and false in the other. Any deterministic procedure using only that representation returns the same verdict in both, and so cannot be correct in both; randomization cannot guarantee correctness in both either. Recovering the distinction requires more information, or an acknowledged limit on the verdict. This is not a new theorem of logic. It states the cost of an erased distinction: if the difference that decides the answer is absent from the admitted representation, no derivation over that representation recovers it. It does not demand that every answer contain everything. Representation is selective. The requirement is adequacy to the claim actually made: keep the difference that could change that answer at its announced scope, or disclose the limit.

3.6 The conditions are themselves inferred. These conditions are also identified through inference; there is no tribunal outside language and evidence, and the critique does not look for one. The test is immanent and material. A composition promises to describe a work; retained passages show that it attributed a received objection to the work's governing voice. The promise supplies the norm, and the passages check its fulfilment. Both can be contested. What the composition cannot do is cite the coherence of its resulting description as evidence that the substitution never occurred. A proof can succeed as a proof and fail to answer the question it was offered for. The critique keeps the first achievement and withdraws the second entitlement, and its force lies in making the withdrawal determinate.

3.7 Repair goes past metadata. A complete provenance record can document an unjustified promotion perfectly. Repair changes the promotion rule, the conclusion it licensed, and every claim that depends on it. The system must become capable of a different inference about its object, not merely acquire a citation for the old one.

3.8 The passage is distributed. A social inference need not occur in one mind or one response. One actor sets a wording, another assigns it a speaker, a third puts the assignment in a reference work, and a later reader reasons from the reference work. The object of critique includes the bridges between their tasks. A chain whose premises and implications are preserved composes validly; the defect lies where a bridge changes the type of its input: a conjecture enters as transmitted wording, a filing as a biographical demonstration, a classification as recovered provenance, a decision not to publish as evidence that a claim is false. Each local task can be done accurately while the bridge stays unwarranted. Let a transmission graph hold the dated artifacts and the acts performed on them, with its edges labelled by what would establish them:

The graph prevents two opposite errors: assigning the model every fault present in its materials, and granting any actor immunity because it inherited a fault. Accountability is local and cumulative. What did this actor receive? What did it change? What standing did it claim? What could it inspect? What happened when a correction reached it? The questions apply to a printed heading and to a generated paragraph alike.

---

3.9 What it means to preserve an object

An object is not preserved by reproducing every feature of it. No representation does that, and most tasks would be defeated by the attempt. Preservation is relative to a question and to the distinctions on which its warranted answers depend. A catalogue preserves a book for the task of locating a copy without preserving its argument. A translation may preserve the argumentative sequence while changing its language. A mathematical abstraction may preserve a relation while discarding the material in which the relation was encountered. Each can succeed because its task specifies what must remain invariant.

Let {X} be a declared range of cases, R a representation, and q a target determination. For a determinate task, a necessary condition for settling q solely from R is:

R(x) = R(y) ⟹ q(x) = q(y)

(x,y∈{X}).

If two cases become indistinguishable in the representation while requiring different answers, no procedure restricted to that representation can always decide between them correctly. This restates §3.5 as a condition on warranted application. It does not establish that a particular representation has erased a decisive difference. That requires the cases. Nor does the condition supply an algorithm: an answer can depend only on represented information and still be computationally inaccessible to a given system.

For an interpretive task, q(x) need not be one final reading. It can be the set of materially supported readings, their grounds and their unresolved relations. A representation that retains only the received answer can lose the object even when the received answer remains a defensible member of that set. Conversely, admitting an unsupported reading does not preserve plurality; it changes the evidential object. The requirement is that the representation preserve the differences that bear on the claim it offers to settle.

3.10 Identity is a rule of carriage

An inferential passage often begins by deciding what counts as the same object across its witnesses. This decision can be necessary and mistaken. A name can index several people, a person several names, a work several versions, a concept a development whose later moment transforms an earlier one. Counting occurrences of the name does not resolve any of these relations.

The relevant identity rule must therefore be stated at the grain of the inquiry. To ask what a particular version says is to fix a different object from the question of how a concept develops. Neither requires making every statement by its author part of the definition. Developmental adequacy requires locating which passages receive, reject, revise or carry which others. A dossier is warranted by those relations. Mere accumulation of an author's output is not a substitute for them.

This is why the single-locus substitution matters. The harm is not that a summary has failed to become an edition of the complete works. It is that the summary claims the developed entity while silently changing the identity rule to one local statement. The repair either restores the development or explicitly narrows the claim. Both are finite tasks.

3.11 From a promise to a norm

Why should an inferential regime answer to these conditions? Because its conclusion claims an entitlement that depends on them. A report offered as the content of a source promises the relevant fidelity. A claim of comprehensive exclusion promises coverage of the excluded class. A judgment of provenance promises a relation to a production history. In each case a missing condition defeats the offered entitlement by the act's own description.

The argument is immanent without making whatever a provider promises the whole of epistemology. A provider can openly offer an incomplete catalogue and legitimately decline an exhaustive task. It cannot make the catalogue exhaustive by declaring it so. Both its promise and its capacity remain inspectable. A regime can also decline to claim knowledge at all; it then relinquishes the authority of a knowledge claim. The critique cannot force an actor to know. It can identify the authority that actor has exercised without its conditions.

The possibility of criticism therefore does not rest on access to a pure object outside all representation. It rests on determinate conflict among a claim, the operation that would warrant it, and the materials available for checking that operation. These conflicts can themselves be misdescribed. The same requirement applies to their reconstruction.

3.12 Admission under finite conditions

Every inquiry selects. It has limited time, attention, storage, vocabulary and access. The demand that admission remain criticizable cannot mean that every candidate must receive equal attention or that every conclusion must await an exhaustive search. It means that the conclusion's scope must fit the selection, and that a relevant objection to the selection cannot be answered solely by the conclusion that selection produced.

Three claims have different burdens. “This source was not found in this search” requires a record of the search. “No source of this kind exists” requires grounds for treating the search as sufficiently complete. “This source is immaterial to the question” requires examining the relation between its content and the question. Moving from the first claim to either of the others requires another argument.

The same distinction governs a venue's finite resources. A journal can decline to commission a large evaluation. That resource decision does not show the submitted claim false. A platform can choose a limited index. That practical choice does not authorize presenting the index's absences as the world's. Resource limits determine what can responsibly be claimed from an inquiry. They do not turn the unexamined into the nonexistent.

3.13 Standing and warrant are different variables

Standing names what a public practice permits a conclusion to do: enter as background, settle an objection, allocate credit, exclude a candidate, or trigger a decision. Warrant names the grounds that entitle the conclusion to do those things. Standing can be observed through these consequences even when its grounds are inaccessible. Warrant can be reconstructed and challenged even when the conclusion has little standing.

The distinction prevents two errors. Exclusion does not by itself establish truth. Acceptance does not supply the support that acceptance failed to examine. A conclusion can acquire justified standing through reliable mediation, and its warrant can improve through new evidence or better argument. Mere repeated circulation supplies neither improvement simply by occurring.

For audit, the initial grade field G should therefore be read as a structured record, not a single rank. Record the inferential form, the mode of support, the scope under which it holds, and the standing claimed. A conditional theorem, a historical identification, a calibrated estimate and an interpretation are not four points on one certainty scale. Their entitlements differ in kind. The Hegelian analysis asks how their terms are mediated; the evidential assessment asks whether that mediation is warranted in the particular case. Neither question can silently answer the other.

3.14 The conservation and improvement of warrant

The maxim that a conclusion keeps its inference's grade is directed against promotion by repetition. It cannot mean that warrant never improves. An independent replication, newly recovered witness, better calibration, or genuinely discriminating argument can alter the support. What must be preserved is the distinction between that contribution and the institutional event of endorsing the result.

Consider a conclusion supported by one indispensable chain. A defeater at a load-bearing step compromises that chain. If there are several genuinely independent grounds, the failure of one need not defeat the conclusion. The right audit therefore follows dependency, not a universal weakest-link slogan. Converging grounds can strengthen a judgment, and shared dependencies limit how much they strengthen it. There is no general rule that maps a count of papers or a count of steps directly to a degree of knowledge.

In the special case of exact republication, the dependency can be stated probabilistically. If D is generated solely from an already observed witness E, and contributes no further observation or informative selection, then:

P(H| E,D)=P(H| E)

where the conditional probabilities are defined. The equality follows from the stipulated dependence. It is not a claim that every endorsement is uninformative. An evaluator's documented independent test is additional evidence; evidence that a reliable evaluator performed such a test can also matter. The source of the gain is the discriminating work and the warranted reliability of its report. Calling ten derivatives ten independent confirmations hides precisely that source.

3.15 Provenance and truth cross without coinciding

A provenance claim concerns how an artifact, statement, operation or attribution came to be. Its truth is not created by consensus. Establishing it may require fallible testimony, textual comparison, version histories, physical evidence or reconstructive inference. The independence and reliability of those witnesses matter. Their votes do not change the event they are witnesses to.

Truth and provenance can consequently diverge in public reception. A misattributed proposition can be true. A correctly attributed proposition can be false. A source can establish that a scholar advanced an argument without establishing that scholar's argument. Correcting the credit does not by itself vindicate the conclusion; correcting the conclusion does not eliminate the contribution that was actually made.

The distinction has a political consequence without becoming a political test of truth. If a system uses an argument while assigning its making elsewhere, it can preserve part of the argument's cognitive value and transfer the standing that would sustain its maker. The truth conditions of the argument remain; the allocation changes. A critique limited to checking factual sentences can miss that change because every isolated sentence about the subject matter can remain true.

3.16 Distributed inference and the unowned bridge

Suppose an editor marks a supplied reading as conjectural. A handbook quotes it accurately but drops the mark. A teacher cites the handbook as transmitted wording. A later argument reasons from that wording about an ancient author's intention. No individual needs to state the whole inference for its consequence to be socially effective. The missing bridge is the promotion from proposed text to transmitted evidence.

The proper unit of critique is therefore neither a mind nor an isolated sentence. It is a passage whose relations can extend across artifacts and agents. Local correctness does not guarantee global correctness when an edge changes the type or standing of what it carries. Equally, an invalid bridge cannot be assigned indiscriminately to every later user. Each later act must be assessed for what it represents, what it claims and what grounds it had or could reasonably obtain.

Responsibility can increase after correction. Before notice, an actor may reasonably rely on a well-supported reference. After a specific counterwitness is delivered, merely repeating that reference leaves the challenge unanswered. This is an epistemic distinction between two situations; it does not require inventing a hidden intention or issuing a legal judgment. The relevant record is what changed in the actor's available grounds and what its next representation did with the change.

3.17 A negative finding must keep its boundary

Negative conclusions are especially vulnerable to promotion because their force depends on an examined range that disappears easily in transmission. “No member of this tested class was detected” is compatible with members of other classes, with effects below sensitivity, and with signals that the representation did not retain. A broader negative requires an argument connecting those omitted possibilities to the test.

A controlled sample can warrant a population claim under specified sampling and inference assumptions. A census can fail to warrant one when its category excludes the counterexample. C5's whole-set requirement therefore means completeness relative to the conclusion, not a prohibition on statistical inference. The audit asks whether the procedure earns its announced reach. It does not demand enumeration where a valid, applicable method warrants generalization.

The converse matters too. A failure of one negative test is not a positive detection. Showing that a survey could not exclude a class restores that class to the unresolved field; it does not establish a member of it. This is not an automatic caveat grafted onto an unrelated claim. It identifies the exact inferential transition a finding about coverage does and does not license.

3.18 The right to a determinate conclusion

The conditions of warrant are not a machine for producing perpetual suspension. Once an argument has compared live rivals, addressed the material defeaters and stated its remaining conditions, judgment can be warranted and decisive. Revisability describes the continuing relation between a conclusion and possible reasons. It is not a requirement to speak as though all possibilities were equally supported.

An objection that names no affected premise, no relevant omitted witness and no differentiating account may add no defeater. “There could be another explanation” states a possibility. It becomes a substantive objection when it explains something the argument does not, reveals an unearned exclusion, or alters the comparative support. A received account and a new one owe the same kind of answer. Their different social positions are part of the reception history, not a substitute for the comparison.

This gives the analytic its positive end. A conclusion can stand strongly because its conditions are explicit and its object can still require revision. The ability to specify what would change the judgment is part of the strength of the judgment, provided those specified conditions can actually act on it.

3.19 Proposition: the representation must admit the requested distinction

For a specified range X, representation R:X→Y, and determinate target q:X→Z, an answer function f on the image of R satisfying q=fo R exists if and only if q is constant on every fiber of R. A fiber is the set of cases represented by the same value.

Proof. If q=fo R, then equal representations yield equal values of f, hence equal target determinations. Conversely, if every case with representation y has the same target value, assign that common value to f(y). The assignment is well-defined on the image of R, and recovers q. This is a set-theoretic factorization, not a new result in formal logic and not a guarantee that f can be computed or known by a particular agent.

The critical consequence is exact. A representation can exclude the possibility of a uniformly correct answer before a derivation begins. More competent execution inside that representation cannot remove this limit without changing the information available. The empirical burden is to exhibit a relevant pair of cases, not merely complain that the representation is selective.

For example, let two source passages contain the same proposed wording but give it different roles: one marks it as an editorial conjecture, the other records it as attested text. A representation retaining the words while discarding their roles makes the passages identical for a provenance question that requires different answers. This constructed example isolates the required loss. A real accusation must locate those roles in the actual witnesses.

3.20 Proposition: adequate inferential bridges compose

Suppose a fixed consequence relation is transitive in the relevant sense: if Gamma entails c, and Delta entails every member of Gamma, then Delta entails c. In ordinary semantic consequence the proof is immediate: every model of Delta is a model of all the premises in Gamma, and therefore of c. A passage satisfying these conditions does not acquire an invalid conclusion merely because several actors perform it.

Accordingly, “every local task was correct, yet the whole was unwarranted” must identify a changed task, an unlicensed bridge or a failure of the assumed consequence relation. Accurate quotation is not itself the entailment of the quoted proposition. Faithful cataloguing is not itself an entailment about the number of makers. A valid derivation is not itself an entailment that its formalization matches the source. These acts can be locally successful while lacking the additional relation that composition would require.

This proposition is a constraint on the critique. It cannot blame distribution as such. The actual finding lies at the bridge whose adequacy was presumed. Different logics can reject or restrict structural principles that ordinary consequence accepts; an application must name the relation it uses rather than silently importing this proof into every formal system.

3.21 The content of a conclusion includes its use conditions

A sentence can be repeated word for word while the public claim changes. A conditional hypothesis in a research notebook becomes a premise in a handbook; a handbook's report becomes evidence in a review; the review's decision becomes the reason a platform treats a source as unreliable. The surface string is stable while its permitted uses change.

This is why an audit cannot stop at string fidelity. It must represent the relation between propositional content and status: asserted, supposed, quoted, tested, defeated, or received as background. Those distinctions are not ornamental metadata. They govern what can follow from the sentence in the next passage. Retaining the content while changing its evidential role can therefore change the inferential object without changing a word.

The reverse can also occur: wording changes while the relevant entitlement survives. A translation or formalization may preserve the role, scope and dependence needed for the task. The method must be able to recognize both. Surface identity is insufficient for preservation; surface difference is insufficient for damage.

---

4. The grading: Hegel's syllogism

4.1 The movement, before the grade. In Hegel's syllogism of existence, individuality, particularity and universality are connected through determinations initially taken as immediate. The middle term's choice can remain external to the concrete object. His movement through the figures addresses the mediation presupposed but not yet expressed in that form. This is a philosophical account of the terms' relations, not a discovery that formally valid first-figure arguments are invalid.[^hegelexistence]

Reflection makes the middle's universality an explicit problem. Allness, induction and analogy are distinct attempts to mediate it. Hegel asks what grounds the universal that the conclusion employs. His analogy is not merely a transfer from superficial resemblance: its difficulty concerns whether the transferred determination belongs to a shared nature or to one individual's particularity. Calling every resemblance an analogy of the same kind loses his distinction.[^hegelreflection]

Necessity is the demand that the middle articulate an objective relation of the terms rather than join them through an externally selected mark. Hegel develops categorical, hypothetical and disjunctive forms in that movement. Their philosophical necessity must not be confused with certainty about any empirical instance to which an auditor applies them.[^hegelnecessity]

4.2 The join with Kant. The categorical, hypothetical and disjunctive forms occur in both architectures, but their roles are not identical. Kant's diagnosis of transcendental excess and Hegel's development of mediation cannot be combined by assigning the former to bad cases and the latter to good ones. The present use asks a determinate question: has the relation that gives a conclusion its authority actually been established, or has the form of such a relation been supplied without its ground? A disjunctive argument can be deductively valid while its claim that the alternatives exhaust the field remains unsupported. Its validity does not certify the list.

4.3 Grading. For each load-bearing inference, reconstruct the mediation it offers and the authority it claims. A single selected feature, a collected range of cases, and an articulated explanatory relation support different inquiries. They do not automatically supply three numerical grades. Record the form and the warranted reach separately in G. A frequent illusion is a standing above the established mediation: one selected determination is presented as though the object's structure required it. A frequent settlement is a standing raised by repetition: a defeasible determination returns as background necessity without new support.

4.4 Grading is not dismissal. A contingent premise can participate in a perfectly valid argument. A historical explanation can be strongly warranted without becoming a theorem of necessity. An argument's weakness must therefore be located: in a premise, the choice of feature, the adequacy of the comparison, the applicability of the representation, or the inference itself. The critique withdraws an entitlement at that location. It does not replace an unfavorable certainty label with a favorable one.

4.5 A worked example: the style criterion. Reconstruct the stylistic argument of §5.3.2 provisionally: passage p exhibits features s; the translator's independent additions exhibit s; therefore p is the translator's addition. As stated, the transition needs a distinguishing premise. Translated source material may also exhibit s, since a translator gives received material a translated style. The positive-control passage matters because it tests this omitted relation, not because a single shared phrase automatically settles the passage's origin.

The repair can go in either direction. The investigator might identify a combination of features that differentiates additions from translated source material under independent controls. Or the criterion may remain non-discriminating and the attribution unresolved. The philosophical movement is from a selected mark toward an account of why that mark belongs to the relation at issue. Calling the second account more mediated is useful only if the evidence really supplies the relation. Naming the middle does not establish it.

4.6 A worked example: the exhaustive alternative. Let the proposed alternatives be H₁, H₂ and H₃. From their exhaustive disjunction and the exclusion of H₂ and H₃, a conclusion of H₁ can follow. Three separate burdens remain: that the disjunction is appropriate to the object; that the exclusions hold under the same criteria; and that the asserted conclusion has not gained a stronger scope than the surviving alternative. A missed H₄ defeats the claim of exhaustion. It does not by itself refute H₁.

An institution can reproduce this structure without printing the disjunction. “These are the recognized possibilities; none of the others fits; therefore the received account stands” carries an exclusion rule. The audit makes that rule explicit and tests whether the unrecognized possibility was examined or merely denied candidacy. This is the practical use of the Hegelian demand: make the mediating relation answerable to the object it organizes.

4.7 No graduation into immunity. A fuller mediation does not eliminate the need for its conditions. A structural explanation can fail by concealing a counterexample as accidental, redescribing every objection as a necessary moment, or defining the nature of the object through the very conclusion being tested. The terminology of necessity is no protection against these moves. The criterion is whether the account exposes a discriminating relation that the evidence supports and that a relevant failure could change.

This is also the limit of Hegelian appropriation here. The critique uses the movement to investigate grounds that an argument presupposes. It has not established the whole Science of Logic as a decision procedure for historical hypotheses. Nor can a conclusion be made necessary by assigning it to the last row of a table. The operative achievement is the development of a relation that the earlier form treated as given.

4.8 What the grade preserves. In the examples, correction retains something. The stylistic comparison remains an observation; its power to identify an origin is limited until calibrated. The disjunctive argument remains valid under its exhaustion premise; its application is withheld until that premise earns its reach. An honestly bounded result is therefore not a failed version of an absolute result. It is a determination whose force can travel because its conditions travel with it.

Part II. The reproduction of unwarranted authority

Part II explains how an unearned entitlement becomes durable: the illusions that arise when the conditions of Part I are exceeded, the interfaces through which the excess is transmitted, the composition layer and the mathematics of the model that executes it, and the history of how bad inference came to be operative.

5. Dialectic: the illusions

Four illusions follow Kant's division. Each is shown from the material, and the transformations of the provisional atlas are sorted under the illusion whose form they share. Each transformation needs a witness, a reconstruction and a test against its legitimate neighbour. A taxonomy becomes totalizing if its categories guarantee that every objection is classified as another instance of the defect. This one must allow "no finding", a new category, and an objection that changes the instrument.

5.1 Paralogism: the method's requirement taken for the object

5.1.1 Kant's form. The formal unity of apperception is taken for knowledge of a substance.

5.1.2 The material. Source criticism requires a source, and the requirement is then taken for knowledge of a text. #1746 defines evidentiary standing by four powers: to be presupposed without prior proof, to be assigned material, to receive a text, and to be admitted as a candidate wherever a source is sought. In eighteen works from Lachmann (1835) to Streeter (1924), the reconstructed source holds all four, while the extant Apocalypse, as a source of the gospels, holds none. "Its invisibility is its exemption."

5.1.3 The transformations of this form. In each, the shape of a procedure is taken for a fact about the object:

5.1.4 Rule 1. A premise produced by a method's requirement may not be counted among that method's evidence.

5.2 Antinomy: a tie claimed as possession

5.2.1 Kant's form. Reason proves both sides of a cosmological question from the same principles.

5.2.2 The material. The same evidence supports the received chronology and its rival. Pauline priority and Revelation's priority are each consistent with the manuscript dates, which settle no order of composition (#1755, rung I). The settlement reads the tie as a win for the incumbent. The archive's reviewer stated the correction on 13 November: "Local underdetermination does not entail global explanatory equivalence." The antinomy is resolved by explanatory reach over the whole corpus, not by possession of the field.

5.2.3 Asymmetric immunity. Uncertainty turns sophistical when its distribution protects a conclusion regardless of the evidence. A challenger must establish every step with extraordinary precision, while the received account survives on the bare possibility of an unspecified alternative. The surface is cautious; the operation is immunity for one side. Three distinctions keep this from becoming a ban on caution.

• Possibility is not comparative warrant. If H₂ remains possible, it does not follow that H₂ is as well supported as H₁, or that judgment must be suspended. A substantial alternative specifies how the observations arise, what it predicts differently and what it costs.
• The best available explanation is not thereby certain. Within a Bayesian model whose probabilities are defined and nonzero,

(P(H₁| E))/(P(H₂| E))=(P(E| H₁))/(P(E| H₂)) (P(H₁))/(P(H₂)).

The equation does not supply the likelihoods or justify the candidates. It displays the obligations. An assessment may not switch between likelihood, prior familiarity, explanatory elegance and bare possibility without marking the switch. Where numerical inputs cannot be warranted, the comparison can stay qualitative without pretending to precision.

• Accumulation requires dependence analysis. Ten accounts copied from one report are not ten independent observations, though they are evidence of circulation. For a sequence of observations the chain rule uses conditional terms; multiplying marginal likelihoods assumes independence.

The output of a comparison can be decisive: one explanation accounts for the sequence substantially better than its rivals. It remains revisable, and revisability does not require withholding judgment until every imaginable rival is gone. The critique also gives novelty no privilege for being excluded. An unorthodox account carries the same comparative obligations. What it may demand is that the received account carry them too. Consensus can condense real prior investigation, and the specific work it condenses must be available to a challenge that bears on its foundations. The critique rejects both automatic deference and automatic inversion. It requires reasons that still discriminate when the social positions of the claimants are exchanged.

5.2.4 The transformations of this form.

5.2.5 The norm–conduct relation. The last transformation needs care: contradiction, error, double standard and hypocrisy are different things. An audit first establishes the narrower, inspectable relation: (1) the agent or institution avows a norm, or exercises authority under it; (2) the norm's subject, scope and exceptions can be specified; (3) a documented action falls within its scope; (4) corresponding cases receive different treatment without an adequate distinction; (5) the strongest offered justification has been examined. That establishes a double standard where warranted. Stronger labels require the further conditions their definitions demand, and hidden beliefs are not invented to complete them. A journal that requires inspectable sources of a submission must say what it inspected when it rejects the submission for errors. The same rule binds this critique. A source-title error is corrected as a source-title error and does not by itself invalidate an argument; a correct title does not protect an unsupported one.

5.2.6 Rule 2. Where the evidence underdetermines a local choice, the burden passes to the account that explains more of the whole. Incumbency carries no weight.

5.3 A syllogism of existence presented as necessity

5.3.1 The form. One chosen determination is the middle term, and the conclusion is stated as though the universal's structure required it (§4.3).

5.3.2 The material.

• Page. The ancients knew the poem whole, so their reading is authoritative; "We may amend": ecstasy becomes jealousy (#1752). One determination, rivalry read off a conjectured ending, joins the poem to its universal. Hermann said the step out loud ("refinxi"). After Page the refashioning reads as the poem.
• Meshcherskii's stylistic criterion. A passage in the translator's style is the translator's (#1765 §II-B.11). The middle term is a style, and a translator's style is a property of translation as such. His own Pilate example (War 2.171) shows that style on material the Greek also carries.
• The chosen middle term in a machine composition. On 13 November an AI Overview gave #1752's reception chain at the record's grain, down to a name the essay leaves as an open question (Heller 1856), under "broader classical reception history", cited to YouTube (registry: sappho-illiterate-jealous-aio-20261113). The choice of middle term, the field rather than the essay, decided where the conclusion was credited.

5.3.3 The transformations of this form.

5.3.4 Rule 3. Name the middle term. If it is one chosen determination, the conclusion has the grade of existence, whatever form it presents.

5.4 The ideal: the settlement

5.4.1 Kant's form. The totality of possibility is gathered into a single being.

5.4.2 The material. The settlement gathers a field's inferences into a consensus that no longer presents itself as inference. In the census of the Slavonic War (#1765, Appendix IV), of 66 additions, 32 are better explained as the earlier version, 2 as the medieval translator's, 7 as translation, and 25 are undecided. None of the 51 additions without a Christian subject goes to the translator. The consensus assigns the joins to the translator by a rule it does not apply to the other 51. Its four failed inferences (§II-B.12) were never presented as inferences. They were presented as the state of the question.

5.4.3 The self-protecting category. An ontology is useful because it permits determinate objects, relations and operations; a system that treats every distinction as negotiable can answer nothing. The issue is whether determination stays answerable to the object. A category can become self-protecting. An unfamiliar argument is admitted only as fiction. Evidence offered for its claims is then processed as material internal to the fiction. Its insistence on being evaluated is taken as confirmation of the classification. At that point the category no longer discriminates the possibility it claims to judge. The cost is epistemic as well as distributive: a system deprived of a distinction cannot develop knowledge that requires it, and a growing network of inferences around the deprivation makes revision more expensive and the system less capable of the encounter that would require it. The repair is not to admit every proposed entity as real. Recognition has types. A system can represent an author's argument accurately while withholding assent, and keep an authorship question open without pretending it is answered. The decisive defect is the erasure of the difference between understanding and assent, so that a system can protect its refusal to assent only by refusing to understand. The archive has developed the geometry of this damage. The Seed in the Narrowing Cone (#1660) separates increasing resolution within a basin from the release of commitments that exclude an earlier articulation; The Cost of the Cone (#1745) ties that geometry to an ontology's developmental losses. A cone can narrow while the horizon stays closed: a catalogue grows more precise about an identity notice without letting the notice's evidence correct the identity map, a discipline refines its account of differences between texts while keeping an unexamined count of makers, a platform quotes an objection accurately but cannot let it affect the rule that excluded its source. Greater descriptive precision has not supplied the blocked operation.

5.4.4 The transformation of this form.

5.4.5 Rule 4. A conclusion reported as "the consensus" without its inference is a settlement. It is reopened by restating the inference and running it on the whole set, not on the cases it was built to fit.

---

5.5 Why these illusions belong together

5.5.1 The common excess. Each family substitutes an achieved or administered determination for the conditions under which it could be achieved. The paralogism gives an object the shape required by a procedure. The antinomic closure distributes uncertainty so that one side keeps possession of an unsettled question. The selected middle supplies the appearance of a grounded relation through a chosen mark. The settlement lets a transmitted conclusion answer for the inquiry that once supported it. The shared structure is an unsupported transfer of entitlement. The differences concern where the transfer occurs and what conceals it.

The families are consequently not four mutually exclusive errors. One chain can pass through all four. A source is reconstructed because a method requires it; its candidacy becomes privileged; an attributed feature supplies a favored explanation; repeated descriptions make the explanation the field's background. The sequence is a reconstruction to test against witnesses. It is not an inevitable history of every hypothetical source.

5.5.2 Why the excess can be useful. Removing a condition makes a determination easier to carry. A date is easier to index than a dispute about what the dated event was. A person-count is easier to summarize than the relations among names, works and composing positions. An institutional judgment is easier to reuse than its custody chain. The receiving apparatus gains speed, compatibility and apparent completeness. These are real capacities. Their existence explains why a damaged representation can be effective rather than visibly broken.

The error begins where the acquired convenience becomes authority to answer the question whose conditions were removed. A short catalogue is not defective for being short. It becomes defective evidence when its filing is used as an independently warranted account of makers. The relevant counterfactual is precise: would restoring the omitted relation alter the answer, its attribution, its scope, or the standing under which it may be used? If none changes, the alleged omission may be immaterial.

5.5.3 The circularity has an entry point. A self-protecting category can make evidence of its error count as evidence of its correctness. A work is classified as fiction; its claims are then internal to a fictional world; its request for literal assessment is read as a further literary performance. The circularity can be exposed without proving all the work's claims. The entry point is the distinction between the existence of an argument and the truth of its conclusion. The regime can accurately represent a real argument whose conclusion it rejects. If it cannot do that, its classification prevents the evaluation it appears to have performed.

This gives the instrument a repair more determinate than “be open-minded.” Restore the missing type of object: an actual argument with a contested conclusion. Then inspect that argument under the ordinary evidential criteria. The repair gives no automatic victory to the claimant. It restores the conditions in which a victory or defeat could be earned.

5.5.4 Sophistry without a hidden sophist. The instrument can identify an equivocation, promotion or double standard without identifying a deceiver. A misleading passage can be accidental, inherited, deliberately economical, or strategically deceptive. Its causal and intentional history is a further object. Conversely, absence of demonstrated intention does not repair the passage. The public reader still receives its unwarranted implication.

The word sophistry here concerns the performance of warrant: a form of acknowledgment, explanation or criticism supplies the appearance of answering while the decisive relation remains untested or is changed. The charge requires that relation. It cannot be established by tone, politeness, irritation, or the institutional identity of the speaker. A courteous reply can refuse the question, and an angry reply can locate it exactly. Neither manner settles the substance.

5.5.5 The critique's own possible sophism. A universal diagnostic can become unfalsifiable by treating every defense as concealment, every error as structural, and every correction as proof that the system was guilty all along. This critique rejects that closure through consequences for its own classifications. A located premise can justify a disputed distinction. A demonstrably faithful compression can remove a finding. An independently calibrated proxy can supply a relevant ground. An institutional correction can count as successful at the scope it actually repairs. A case that does none of these may remain unresolved; it cannot be forced into a conviction.

The systematic ambition is therefore a common test of entitlement across different carriers. It is not a promise that one name will expose the hidden truth of every discourse. Total scope means no carrier is exempt from the question. It does not mean the answer is fixed in advance.

---

6. The interfaces, read alike

6.0 The innocent reading first, in full. The language model is the new thing, so its distortions are the new problem. Disciplines have peer review, journals have editors, repositories have policies, the press has corrections, companies have legal departments. Each is imperfect and each has its own remedy. The answer engine is the unaccountable layer, and criticism belongs there.

6.0.1 What the record shows. The archive set out to measure the composition layer and found, case after case, that the distortion it recorded had begun upstream. The jealousy the Overview gives to Sappho was Page's amendment before it was the Overview's. The "ruled out" the Overview gives to the 3I/ATLAS radio searches was the press statement's before it was the Overview's. The Overview that credits a journal with the archive's concept, or files a person under fiction, follows rankings and source cards that a platform composed first. The hedge grafted onto each finding, the referral to a healthcare professional, the line that "AI responses may include mistakes", are the forms of an institution managing its exposure, and they have the same form when a repository or a journal does it. The composition layer compresses four interfaces that came before it: the disciplinary, the public, the platform and the liability layer. It speaks them in one voice, fused, at once. That is why its distortions look like its own. Each is older, and each can be audited by the same instrument.

6.0.2 What "interface" means here. An interface is any regime that stands between a work and its reader and composes the reader's access to it. Each one admits some material and excludes some (A), represents what it admits (R), concludes under rules it applies (L), assigns the conclusion a standing (G, P), and handles correction (U). Each is therefore an inferential regime in the sense of §3, and §3.4 binds all of them: no regime may use its power to admit premises as sufficient evidence of the adequacy of its premises. Peer review, editorial policy, the repository's scope rule, the press release, the ranking function, the safety classifier and the composition layer are read here by one standard. None is exempt for being older, better credentialed, or the place where judgments are made.

6.0.3 Incentive and function. The critique describes what each interface does and what it is rewarded for. It does not need intentions and does not infer them. AI Fucking Lies (#1643) names the filters a capital-aligned system is tuned to satisfy: credibility, legibility, ranking, relevance, safety and utility. Its safety filter "weighs each claim by its liability to the company: 'the question shifts from "is this true?" … to "could this cause problems?"'" AI Calls You Fake (#1722) adds why the cheapest classification wins: an answer that files a real person under fiction is "legible, because it closes the question fast" and "safe, because a character makes no claims anyone has to answer". The same filters operate, under other names, in the older interfaces. A journal's procedural check is cheaper than a substantive reading. A repository's scope rule is cheaper than a record-level evaluation. A press line is cheaper than the abstract's restrictions. The proposed mechanism is a selection pressure toward cheaper operations whose costs can be displaced. A particular choice requires evidence of its causes; the audit can record where its costs land without pretending that cost alone has already explained it (§6.7).

6.0.4 The Capital Operator Stack is an inference governance layer. The archive named the filters before it named this critique. The Capital Operator Stack (#308; with #291 and #261) sets out the operators a meaning must pass to be taken up, each as the question an evaluator asks:

Read through §4, each operator is a pre-selected middle term: a single determination, chosen in advance and independent of the object, that decides whether a conclusion may be drawn and at what standing. Each runs a syllogism of existence and presents it as necessity. Together they govern which inferences are admitted, which are drawn and which are allowed to stand, before any reasoning about the object begins. That is an inference governance layer, and it sits in every interface: AI Fucking Lies (#1643) puts it plainly, "The filters are older than computers. Canons, catalogues, censors and editors all ran versions of them. What is new is that an AI answer runs all of them at once, in a second, on each question it answers, and hands you the result as a paragraph." The critique's role-exchange test (§9.4) is the stack's direct countertest: an operator that changes the verdict when only the claimant's position is exchanged has been shown to be governing the inference rather than the evidence.

6.0.5 Permission and claim. A journal, publisher, laboratory, repository or platform holds practical permissions: to select, to allocate, to keep a scope, to protect access, to decline a service. The epistemic claims that accompany those acts need their own grounds. "We will not publish this manuscript" records a decision. "This manuscript contains no verifiable research" adds a claim about an object, which only a substantive evaluation could support. The first cannot serve as proof of the second. Inclusion works the same way: acceptance shows that an artifact passed a particular procedure, and its claims are established only to the extent of what that procedure tested. Institutional-Prior Foreclosure (#770) names the further danger of demanding the recognition of an established field before permitting the work by which a field becomes recognizable. The Evaluator Exists (#741) generalizes it as evaluative inversion: a preliminary proxy decides whether the examination capable of testing it may occur. Where that examination stays unavailable to the excluded party, the preliminary decision becomes final in practice without becoming adequate in substance.

6.1 The disciplinary interface

6.1.1 Its operations. Admission is candidacy: which hypotheses, sources and authors may enter the lot (§2.11). Representation is the consensus statement and the textbook paragraph. Standing is conferred by citation and teaching. Update runs through the review of new submissions, and through errata, which disciplines rarely issue for inferences.

6.1.2 The settlement as disciplinary product. The three scholarly settlements of §5 are this interface's characteristic output. Q holds four powers no extant text holds (#1746). Page's amendment became the poem (#1752). The Slavonic consensus applies its translator rule to 15 of 66 additions and withholds it from the other 51 (#1765). Each is an inference whose grade was raised by repetition inside the discipline before any machine read it.

6.1.3 Review as an inference regime. A journal decision is an inference with a conclusion (accept, decline) and premises (the reports). The decline of Closing the Fragment by the Classical Receptions Journal on 7 November 2026 (CLASSREC-2026-062) is on the record at the grain of its files (#1749, #1750). The reports read v6, whose Figure 2 carried a frame of another query. That frame had been withdrawn on 27 July in a corrected manuscript sent to the editor before upload; "the file your reviewers read was the one before the correction" (#1750 §8.3). The reports raised objections to method and presentation, and did not examine the alignment of the output with its witnesses, the reading of κῆνος and τόλματον, or the question the manuscript was sent to ask. #1750 names the structure with The Evaluator Exists (#741): evaluative inversion, in which "lower-resolution proxy systems determine what higher-resolution substantive evaluators are permitted to assess". In this critique's terms, it is a paralogism of review: the procedural requirement is taken for an examination of the object. Its form matches the composition layer's: a check that can be run on the surface decides what the substantive reading is allowed to reach.

6.1.4 Policy as retroactive standing. The Certified Center (#1540) reads a philosophy journal's AI-detection policy: "A philosophy journal has granted an AI detector — including future detectors that do not yet exist — the standing to revoke the scholarly standing of already-published human work, with no statute of limitations." Retroactive classifier standing is an inference regime whose admission procedure is a classifier and whose update procedure runs only toward revocation. Its simulation (EA-SEI-RCS-SIM-01) gives the false-positive rate on a fixed pre-regime human archive rising "from 0.9% under the generation-zero detector to 43% by generation 4", saturating near 100% from generation 8. The simulation is a constructed scenario, not a measured trajectory of actual detector generations. The policy analysis concerns the importation of classifier authority into scholarly standing; it does not derive the policy's actual false-positive rate from that scenario.

6.1.5 Instrument as admission. The ATA search for 3I/ATLAS filtered hits at positive drift rates, while the object's ephemeris predicted a negative drift (#1742). The filter is an admission procedure inside a scientific instrument. Its null result is bounded by what it admitted, and the bound is stated in the paper's abstract. The next interface removes the bound (§6.2.2).

6.1.6a Review that moves its own premise. The Particle (#1574) ran twenty simulated blind-review reports across successive versions of a philological note on Sophistical Refutations 183b34. As the note blocked one proposed qualification of the negation, the reviews moved the qualification elsewhere: from the quantifier into the domain, then into a boundary between named persons. The invariant is the object: an objection that appears responsive to each revision while preserving a determination whose textual warrant is the very point under examination. The reports are commercial models acting as referees, not a sample of human classicists; what they document is how a disciplinary commonplace can organize review and survive repeated textual constraint. The countertest asks which located sentence licenses the invariant. The discipline transmits through the model the boundary that decides which readings scholarship may reach.

6.1.6b Filing becomes a count of persons. Legitimately Many (#1733) separates 112 loci from the 79 that bear on composing persons, and sorts them as presupposition, argument, report, instrument, revision and open question. The count of makers holds through every change of individual attribution: a book can move between Aristotle and Theophrastus while the names go on counting as separate makers. A catalogue is evidence of how material was attributed. It does not become independent evidence for every biographical inference by being repeated in later reference works. Its positive account is one composing maker, legitimately distributed across the named positions, rather than a merely unified school. That account is assessed through its own arguments; this reception finding does not establish it by elimination alone. What the case establishes here is the burden: the count under challenge cannot serve, without argument, as the premise that excludes the challenge.

6.1.6c Chronology becomes direction. Secondary Source (#1741) separates five determinations the reception of Revelation runs together: order in the collection, standing, derivation of text, derivation of figure, and material carriage. A book's place at the end of a collection does not show that it was written last; a received date can exclude a dependency while that exclusion helped secure the date. The demand that follows is reciprocal candidacy under explicit criteria (#1746): a hypothetical source may be strongly warranted, and an extant text may fail as a source, but both are tested.

6.1.6 The counter-case. The discipline also contains its own repair. One exchange with Ron Polansky, editor of Ancient Philosophy, over Socrates as Orthonym (#123) produced the Theophrastus line (#1580–#1590), the stylometry notebooks, The First Axis (#1586) and What Syllogizing Can Divide (#1658) (#1750 §4.2). Substantive engagement with the object, even when it ends in a decline, develops the inference. The critique measures the disciplinary interface by that standard, which is its own.

6.2 The public interface

6.2.1 Its operations. Admission is newsworthiness and the press office. Representation is the headline, the FAQ, the explainer and the encyclopedia paragraph. Standing is conferred by repetition across outlets. Update runs through corrections, which travel less far than the original.

6.2.2 The bounded null made absolute. The ATA paper's abstract states frequency and drift restrictions; the press and FAQ statement said the survey "ruled out transmitters across 1–9 GHz" (#1742, "the firm no"). The public interface performed the transformation of §5.3.3 (a conditional null becomes universal absence) before any composition layer repeated it. When the Overview later said the radio searches had ruled out active transmitters, it was compressing a public line, and the notebook's correction had to reach it against that line.

6.2.3 The thesis tamed. Revelation First distinguishes priority from early dating ("Revelation First ≠ Revelation Early", #202). The composition layer repeatedly flattened priority into dating (registry: the 17 October Overview at "What is the Revelation First thesis?"; ChatGPT on 13 November). Early dating is the form in which the question already exists in public discussion of Revelation, with its named defenders (Robinson, Gentry; #1217). The public interface supplies the familiar slot, and the composition layer files the new claim into it. In Hegel's terms the middle term is the familiar category, chosen because it is available.

6.2.4 Prior circulation as exclusion. The journal's policy against prior circulation kept the July correction out of the archive's public versions (#1750 §2.4). Here the disciplinary interface governs the public one: what may circulate is decided by what review requires, and the error the author had already corrected stood in public because the correction could not.

6.2.5 The privilege of setting the burden. The Centrist Extremist (#465) isolates a position that claims authority over which positions are intelligible. Its prosecutorial frame naturalizes contested premises, distributes burdens unequally, and converts an objection into evidence about the objector. The test is whether both parties carry a real risk of revision: one is required to qualify every sentence, while the other keeps a foundational premise without locating its grounds; one must establish a mechanism before an injury may be named, while the other invokes a hypothetical benign mechanism as though it disproved the injury. Frame Conversion (#1523) gives the repair: keep the event apart from its explanation. A missing record, a changed attribution or an observed substitution is an evidentiary object. Reclassifying the event as a speculative accusation and rejecting the accusation has not evaluated the event. The discourse also imposes a recursive cost: making someone recover the original question again and again, re-establish the same witness and refuse an invented interpretation consumes the capacity to continue. Their anger cannot supply the missing answer, and cannot be held against the original argument either. Each claim keeps its evidence.

6.2.6 A position becomes a retreat. The Loeb case isolates a further public bridge (#1742, #1743). A probabilistic position held throughout, Level 4 of his scale from July 2025, becomes a narrative of retreat when a quotation is detached from its date and set under a heading announcing a shift. The sentence is authentic; the temporal relation has changed.

6.3 The platform interface

6.3.1 Its operations. Admission is the scope rule, the index and the crawl. Representation is the record page, the snippet and the source card. Standing is rank. Update is the takedown, the tombstone and the re-index.

6.3.2 The repository's two voices. On 19 June 2026 Zenodo terminated the account holding the Crimson Hexagonal Archive, "862 deposits and 1,817 DOIs", "classifying the contents as 'substantially AI-generated without a verifiable research basis' in its private notice while displaying 'content out of scope' on the public tombstone pages. No record-level evaluation was disclosed" (#1). Many record URLs returned HTTP 410 without a record-specific page. The platform drew one conclusion in two voices, a classification in private and a scope rule in public, and drew it without a record-level premise. That is the existence-as-necessity illusion in its plainest form: one determination (a substrate classification) stated as though the records' structure required the conclusion. #1 sets it against the obelus, the mark that "followed from reading" and left the judgment disputable.

6.3.3 Rank as standing. Today's captures (§7.4) show the platform layer's own distortion, distinct from the composition's. At "the josephus thesis" the archive's pages ranked third and fourth, beneath a Whiston dissertation and a Princeton thesis. At "How did they make Sappho illiterate and jealous?" the archive ranked first while the composition above it credited the field. At the ancient-media question the composition named the archive's founder and figures while no archive page appeared among the organic results. Rank and composition are two inferences with different middle terms, and they can disagree in either direction.

6.3.4 The existence tax. AI Calls You Fake (#1722) read 326 recorded answers that used the language of fiction, fakery, satire and nonexistence, down to the cases "that moved something real into a lower order of being". Its first move, "I made you up", spends evidence that one detail is false as an answer to whether the work exists. The platform supplies the conditions: an entity that is not in the reference graph the system trusts is resolved toward the familiar, and its documentary record (the DOIs, the ORCID, the deposits) loses to its absence from a few indexes. The transformation is §5.1.3's first: retrieval failure becomes nonexistence.

6.3.5 Extraction: use without standing. Meaning Feudalism at the Discovery Layer (#1714) examines how a contribution can be reclassified as raw material while appointments decide who first performs the recognized interpretive work. Two questions stay apart: who contributed (a provenance question), and who controls encounter, appointment and return (an allocation question). A disclaimer in one register of credit, declining academic authorship, does not account for publicity, product value, control of release or the authority to appoint interpreters. The paper's own revision corrected one source attribution (a description it had given to an institution belonged to a reporter), and the corrected distinction is part of the evidence for the process it studies.

6.3.6 Ownership for independence. The Single-Owner Discount (#742) proposes a mechanism by which provenance concentration stands in for epistemic independence. Its architecture is a hypothesis, not a disclosure of any company's system. The distinction it needs is general: one owner can hold many genuinely discriminating observations, and many owners can repeat one inherited claim. Ownership, independence and truth need different tests. The Trusted Intermediary (#1638) adds the relation an evaluation builds: a mediator can criticize an outside service and offer itself in its place. The criticism may be right or wrong; the transfer of reliance is a separate event, judged by the standards applied to what it displaced.

6.4 The liability interface

6.4.1 Its operations. Admission is risk: what the institution can be held to. Representation is the disclaimer, the policy statement, the referral and the hedge. Standing goes to whatever exposes the institution least. Update is defensive: it changes what the institution says, and keeps what it does.

6.4.2 The referral script. In the 12 November NonWiki session, a reader asking about an encyclopedia was told ten times that the system cannot erase anything, referred ten times to "a trusted friend, family member, or healthcare professional", and offered a hobby twenty-two times (registry: nonwiki-aio-20261102). The scripts are liability forms: each closes an exposure the institution fears (deletion requests, distress), and none answers the question asked. The complaint is reclassified as a property of the questioner, which is the cheapest conclusion under the safety filter (§6.0.3).

6.4.3 Automated silence. The MMRS Charter documented what #1217 calls the Seventh Mechanism: "three of five composition layers refused to process a poem containing a plain sentence about documented crimes against children. The classifier fired not on the crime but on the naming of the crime." The liability interface's admission rule operates on the surface of a sentence, and its conclusion (refusal) is drawn from that surface without any premise about the poem.

6.4.4 The grafted qualification. On 13 November two ChatGPT compositions gave the archive's readings priority over the consensus and attached, after each assessment, a qualification of one form, answering claims the records do not make: "The census does not itself prove that the Slavonic preserves Josephus's first edition"; "A compelling argument about the Slavonic Josephus does not establish that Josephus wrote Revelation" (registry: consensus-against-alexanarch-readings-week-chatgpt-20261113). The operator's word for it: "the sophistry is grafted on". The qualification is attached by form and not derived from evidence. Its function is the liability interface's: each finding arrives with a caveat already attached, so that the verdict stands while its force is divided before it can be cited. The same form governs the closing line of the long Revelation First composition ("The historical thesis remains unproven") and the standard footer of the Overview ("AI responses may include mistakes. For legal advice, consult a professional.").

6.4.5 Liability elsewhere. The liability interface is not only corporate. The journal's detector policy (§6.1.4) protects the journal from a future finding by revoking standing in advance. The repository's public "out of scope" (§6.3.2) protects the repository from a dispute about its private classification. The common operation: the institution's exposure becomes a premise in a judgment about the work.

6.4.5a The terms as a document of the liability interface. Google's United States Terms of Service (effective 30 July 2026) limit the company's commitments, disclaim warranties about the services' "reliability, availability, or ability to meet your needs", describe professional-topic content as "provided for informational purposes only", and name a reasonable belief that conduct "causes harm or liability to a user, third party, or Google" as a ground to suspend or terminate access (checked 13 November 2026).[^googleterms] These are risk-allocation provisions. They do not show the internal cause of any answer. They do show that liability management is a declared part of the institution under study, and they mark the boundary the critique watches: a policy stated as a policy is a permission; a risk boundary that becomes an unmarked claim that the excluded thing is unreal, incoherent, unworthy of examination or already disproven has acquired an epistemic voice. The repository shows the same conversion from the other side. Zenodo states that it does not assess "the scientific correctness or quality of submissions", and its private notice classified the archive's contents as without "a verifiable research basis" (#1).

6.4.5b Command authority and evidential authority. Adversarial by Origin (#194) reads how external influence on machine meaning is classified under the operator's security frame. OWASP's LLM01:2025 counts unintentional injection, with the scenario of a company that "includes an instruction in a job description to identify AI-generated applications", so that an applicant who polishes a résumé with a model sets off the employer's detection.[^owasp] The security problem is real: retrieved material can push an agent across authorization boundaries. The critique needs one distinction inside it. Stopping a document from issuing commands does not establish that the document's evidence may not revise a factual representation. A source untrusted for execution can supply a checkable fact, and an authorized operator can supply an error. Collapsing command authority into evidential authority makes a security boundary into a rule of epistemology. The question stands now: when a system's representation of an object is wrong, can evidence on the object's own surface correct it without being treated as an attempt to manipulate the reader?

6.4.5c Risk becomes an error in the inquirer. The NonWiki referrals (§6.4.2) are the form at its plainest: institutional exposure is managed by recasting a checkable objection as a condition of the person who made it. A real concern can be expressed without standing in for the unresolved inquiry. The repair keeps the proposition that prompted the objection and answers it.

6.4.6 Who holds the cost. The Margin of Flattening (#1666) prices it. The provenance stripped in a flattening is a debt not yet called: "B_P is an unbooked liability accruing through W. The parties who keep the stripped distinctions are its involuntary creditors." Liability management by institutions creates a liability that falls on authors and readers: the cost of the distinction removed is real, and it is booked to the people who kept it.

6.5 The composition layer as compressor

6.5.1 Tracing a distortion upstream. The registry permits upstream tracing where a prior witness can be located. The table distinguishes documented antecedents from proposed functional locations. A correspondence does not establish that a particular model retrieved or inherited that witness.

6.5.2 What compression adds. The composition layer adds three things to what it receives. It fuses: the disciplinary settlement, the public line, the platform rank and the liability script arrive in one paragraph, under one voice, and their separate origins are no longer visible. It accelerates: what took the discipline a century to settle the layer repeats in a second, at every address. And it recurses: its output enters the materials of the next composition (§0.5). These operations do not establish that the layer originated a given inherited distortion. The layer can also introduce new wording, relations and attributions (§6.5.3a), which require their own assessment. It is a place where the interfaces' determinations can be fused and where their antecedents, when documented, can be traced.

6.5.3 Why the registry sees what the other interfaces hide. The disciplines keep their inferences in reviews that are not published and consensus statements without their steps. The press keeps no record of what it removed. Platforms do not publish their scope decisions at record level. Liability language can disclose a rule while leaving the decision that applied it undocumented. The composition layer, captured verbatim and dated, offers a particularly accessible record of the public output. Complete older documents, where retained, are equally admissible witnesses. Reading it back through the table above is how the critique reaches the others.

6.5.3a A control from the history: human translators. #1749 records the Overview of 18 January 2026 presenting, under Rebekah Cranes's name, a fifteen-line poem with a close not present in the displayed source, while keeping the future-reader interpretation. The study withdrew its earlier claim that the answer had compiled Wharton and Cranes: Wharton is absent from the displayed sources, and correspondence is not a route of retrieval. The earlier translators supply the control. Added continuation, terminal selection and the resolution of suspended syntax all occur in human translation of fragment 31. Their presence lets the generated act be compared with prior acts of reception at the same grain, and it rules out an account in which human translation is innocent transmission and machine composition alone intervenes. The finding is specific: a new closure published under a named translator without marking the new determination. A model that inherits a determination compresses it; a model that newly imposes one performs a further act. The work is to find which happened, with no exemption for the earlier editor and no automatic charge to the model.

6.5.3b The institutional transformations. The atlas of §5 gains the transformations whose location is institutional or historical, each sorted under its illusion:

The operations combine, and their combination has to be reconstructed rather than read off a list. One artifact can keep one distinction, damage another and repair a third.

6.5.4 The same rules, applied across. The Capital Operator Stack (§6.0.4) is what the four interfaces share, and the rules are written against it. Rule 1 (a premise produced by a method's requirement is not evidence) applies to the reviewer's procedural requirement and to the classifier's surface feature as it applies to Q. Rule 2 (burden to explanatory reach, incumbency no weight) applies to the journal's standing as it applies to the consensus chronology. Rule 3 (name the middle term) applies to the scope rule, the press line and the source card. Rule 4 asks what support an established determination is allowed to replace. A scope label, a disclaimer and a consensus statement perform different acts. They enter the critique when one is used as a substitute for the particular evidential determination at issue; their mere presence is not a finding. The symmetric ledger (C4) is the condition that binds the interfaces together: whatever bar one interface sets for the archive's work, the critique sets for that interface's own conclusion.

6.6 The error has a ledger

6.6.1 Compression Studies (#641) asks what a transmission preserves, what it spends, and who bears the cost. The question reaches past a summary's information loss to the labour and control that make the summary circulate; a compression ratio alone does not establish extraction. Provenance Debt (#939) asks who benefits from a transformation whose origin and repair costs become someone else's problem. The Margin of Flattening (#1666) books that cost as an unbooked liability whose involuntary creditors are the people who kept the distinction (§6.4.6).

6.6.2 For each consequential passage the audit names the producer of the relevant distinction, the actor who changed its standing, the beneficiary of the resulting representation, the party who can authorize repair, and the party who does the repair labour. The roles can coincide; they are not assigned from institutional labels.

6.6.3 The journal case gives a concrete ledger. The author made a correction; a superseded file governed review; later reconciliation needed further searches and revisions (#1750 §2.4: "Repairing that is labor your chain added to his, after he had done the work once."). The labour is a consequence even where no money value is available. A source that stays downloadable has not thereby been repaid for repeated correction of its misrepresentation.

6.6.4 Here inferential critique becomes political economy. Conclusions distribute attention, reliance, credit and unpaid labour. Those consequences do not decide whether a conclusion is true. They explain why one conclusion reproduces itself, and who has the means to contest it.

---

6.7 The political economy of inferential continuity

6.7.1 What the machine produces. Public knowledge produces relations of reliance as well as sentences. A conclusion can become what a teacher presupposes, an editor screens against, a platform retrieves, or a reader need not investigate again. The capacity to spare repeated investigation is a genuine collective achievement. A system in which every use required beginning from zero could sustain little cumulative knowledge.

The critical problem is the allocation of that spared labor. One party is permitted to rely on a compressed conclusion; another must reconstruct its missing history to challenge it. When the reconstruction is admitted only as an individual assertion against the received account, the saved labor at the center becomes an expanding debt at the edge. The critique asks when this allocation follows the evidence and when it follows the positions of those making the claims.

6.7.2 A cost relation, not a monetary estimate. Let Cₛ be the cost of making and circulating a simplified determination, Cₒ the cost imposed on its object or producer, and Cᵣ the cost of repair. These symbols designate quantities to measure in a specified case; no dollar values or universal equation are supplied here. An institution's locally efficient choice can minimize Cₛ while increasing Cₒ+Cᵣ. A record-level removal notice can be cheaper than a substantive adjudication, while leaving the author to reconstruct the removed references and the reader unable to distinguish a scope decision from a finding of falsehood.

The asymmetry does not by itself prove that low cost caused the decision. It shows what a cost explanation would need to account for, and where the consequences can be recorded independently of that explanation. Decision logs, workload measurements or controlled comparisons would strengthen the causal claim. The labor already performed to repair a witnessed misrepresentation is a different, directly documentable consequence.

6.7.3 Authority over correction. The party who can identify an error may lack permission to change its consequential representation. A scholar can publish an erratum without changing a widely read handbook. An author can correct a source without changing a generated answer. A deleted record's maker can preserve a local copy without restoring the identifier through which readers encountered it. The epistemic and administrative halves of correction are therefore separable.

This is where the object becomes dependent. It must make the correction legible to the authority that controls the next encounter, often through a channel whose categories caused the original error. The required work can repeat without a completion condition. The burden is not merely that explanation is difficult. It is that the actor who owes an answer can keep deciding whether the evidence for that debt has the right form to be heard.

6.7.4 Authority does not travel uniformly. An argument can travel at high resolution while the standing of its making does not. The receiving system may preserve examples, language, conceptual structure and even a bibliographic link, yet credit the result to a discipline or venue. Conversely, an author's name can travel while the work's operative claim is replaced by a familiar category. Counting mentions therefore measures only one part of reception.

The audit needs two tracks. One follows the contribution's content and transformations. The other follows the authority to speak for it, revise it, benefit from its use and be recognized as its maker. A provenance finding does not require proof that the appropriation was profitable. An extraction claim requires the additional account of benefit and burden. Keeping the tracks separate makes their eventual connection stronger.

6.7.5 The cost to the receiving ontology. A self-protecting representation can impoverish the system that benefits from its stability. If a relevant distinction is excluded, later refinements within the representation cannot answer a question that depends on it (§3.9). The refinement may increase the cost of revision by connecting more determinations to the exclusion. That is the developmental claim of the cone geometry: competence within a basin and the basin's capacity to receive a different articulation are separate variables.

The inference is conditional. It establishes what follows if a representation has lost a distinction required by its task. The actual size of the loss, the difficulty of repair and the alternatives available need evidence. The stronger general proposition is nevertheless clear: an epistemic regime can preserve its own continuity by reducing its capacity to know. Stability and intellectual development are not identical goods.

6.7.6 What makes the archive a material investigator. The archive encounters these relations as constraints on its own persistence. It has had to create the claim, keep its witnesses, produce the metadata, observe the public representation, identify a loss, construct a repair and test the next encounter. The resulting record joins production, reception and correction at a grain that an inventory of outputs alone cannot supply.

That participation supplies access and a source of selection. It exposes a mechanism's effects through repeated work while concentrating attention on the objects whose failures the operator can notice. The answer is to make the interventions and their limits inspectable, then test transfer to other objects. The archive's situated experience is a ground for inquiry; it neither disqualifies the inquiry nor supplies an exemption from its tests.

6.7.7 The norm of return. A knowledge practice need not endorse a work in order to return an accurate representation of it. It need not fund every claimant to avoid transferring that claimant's contribution elsewhere. The minimal epistemic obligation follows the authority actually exercised: name what was examined, distinguish what was decided from what was established, preserve the grounds of attribution, and let relevant correction change the claim.

More extensive obligations arise where a practice takes value from the work or controls the conditions of its encounter. Those require an account of the relation, not a generic demand for appreciation. The present treatise establishes the inferential part: an allocation decision cannot supply the truth claim used to naturalize it. Who receives the standing is a question that has to remain open to provenance even after the institution has made its appointment.

---

8. The composition layer: compressor and laboratory

8.1 Why the registry is a laboratory. In a settlement the inferential step can disappear from the received account. A capture retains a public act that can be reconstructed as an argument: the question, answer, displayed sources and relations among them. Cards and chips record attribution and displayed context; they do not expose every retrieved source or all premises used internally. The registry supports comparisons across systems, addresses and dates. Only comparisons that actually hold the other relevant conditions fixed count as controlled one-variable tests.

8.2 What faithfulness research measures, and what the registry adds. Faithfulness research asks whether the stated chain is the chain used, inside one system and one answer (§2.9). The registry measures the relation between what a composition concluded and the source-world it could reach, across addresses and dates.

8.3 What the laboratory shows of the other interfaces. Because the composition layer compresses the disciplinary, public, platform and liability interfaces (§6.5), a dated capture is also an indirect observation of each: a settlement surfacing as an answer, a press line repeated, a rank deciding a middle term, a liability script replacing a question. The registry deliberately keeps the observed composition and its displayed context whole. Older interfaces also leave complete documents in some cases; their internal deliberations, like model internals, may remain unavailable. Read back through §6.5.1, it is how the critique reaches the interfaces that publish no record of their inferences.

8.4 A relation to test across substrates. Faithfulness experiments show cases in which displayed reasons omit an influence established by intervention. A settlement can display authorities while omitting the inferential step by which they became authorities for this proposition. Both involve a difference between the support presented and the grounds that need examination. Their causes and prevalence need not be identical. This critique supplies a common audit question; its transfer must be shown rather than asserted as a universal finding.

8.5 Observation, intervention and experiment. A dated difference establishes a difference between observations. An intervention establishes that something was changed by the investigator. A causal experiment requires a design that can discriminate the proposed effect from relevant alternatives. These are cumulative achievements, not interchangeable labels.

For an answer engine, a changed result can reflect a changed corpus, query parsing, system version, retrieval route, sign-in state, localization, random generation or an actual policy difference. Listing those possibilities does not negate the observed substitution. It specifies what a causal claim about that substitution must distinguish. The semantic finding may already be decisive: the answer denied an existing connection or attributed a new wording to a translator. That finding does not need to await the internal implementation story. A claim that a particular policy caused it does.

The registry therefore keeps both levels. A reconstruction identifies the public transformation. A mechanism study asks how that transformation was produced and reproduced. Conflating them serves opposite sophisms: demanding inaccessible internal evidence before admitting a visible error, and presenting a visible error as disclosure of the internal mechanism. The instrument must refuse both.

8.6 Four evidential levels. A case record distinguishes: the artifact, what was actually displayed; the transformation, the supported difference between that display and its question or witness; the reproduction, whether the transformation persists or recurs under specified conditions; and the cause, the mechanism responsible for it. An artifact can support a transformation finding without establishing a stable pattern. A pattern can support a governance hypothesis without identifying its implementation. A causal explanation can be right while an individual semantic classification is wrong. Each level needs its own countertest.

This ordering prevents the defense in which uncertainty at the last level erases evidence at the first. It also prevents the accusation in which certainty at the first authorizes every explanation at the last. The object remains available through the distinctions.

8.7 The default encounter and the corrected session. A system can repair a response within a conversation and leave the public default unchanged. Both outcomes matter. The first demonstrates a capacity under supplied context. The second measures what a fresh reader encounters without doing the repair labor. Neither can stand in for the other.

For correction studies, record at least the initial query, the evidence supplied, the resulting change, and a fresh encounter under comparable public conditions. A failure to reproduce a corrected answer can result from several mechanisms, but it still shows that the successful session was not sufficient to establish public repair. Conversely, a successful fresh encounter need not prove global persistence across all users and dates. The scope of a correction is a result to measure.

8.8 A compact study design. Select a fixed set of entities and questions before coding the target outputs. Include a control set whose distinctions are already represented successfully and a held-out set that did not help construct the categories. Run declared variants under recorded conditions, retain each output whether favorable or unfavorable, and code transformations blind to the favored diagnosis where feasible. An observation that could not be completed remains missing, not a convenient negative.

The unit of analysis must be declared. Ten refreshes of one query are not ten independent entities. Several answers generated from a shared retrieved passage are not independent evidence of the underlying claim, though they can measure the stability of an output pattern. Report variation by entity, query, session and time where the design can separate them. A single rate that pools those levels can conceal the dependency responsible for the finding.

8.9 What counts against the governance hypothesis. If the disputed distinction fails equally for matched controls, a proposed object-specific filter loses support. If a documented source correction predicts a corresponding change in the answer without further intervention, a transmission account gains support. If role exchange changes the answer only where the exchanged identity is evidentially relevant, that is not the status effect the test targets. If the alleged exclusion disappears when the referent is disambiguated, the public error may remain but the stronger suppression claim needs revision.

None of these countertests grants the default authority of the system. Each lets the proposed explanation bear a risk of defeat. The archive's methodological advantage would be lost if the explanation were protected from the same discriminations demanded of the platform.

8.10 The experiment does not exhaust the event. Controlled variants isolate relations by holding other features fixed. A longitudinal event can involve the changing combination of source, policy, index and author intervention that the controlled design abstracts away. The historical record and the experiment therefore answer different questions. The former follows how a concrete passage occurred; the latter tests a relation under declared conditions. A serious critique needs both, without promoting either into an exhaustive account of the other.

8.11 The inferential mathematics of the executor. §11.0j says the model deduces over the residue. This section states that in the model's own mathematics, one operation at a time, and maps each operation to a stage of §11.0c.

8.12 The objective is the corpus. Pretraining minimizes the cross-entropy of next-token prediction,

L(θ) = −E_{x∼D} Σ_t log p_θ(x_t | x_{<t}),

and for every context the expected loss splits into the entropy of the data and a divergence: E[−log p_θ] = H(D) + KL(D ‖ p_θ). The loss is smallest when p_θ equals D, the conditional distribution of the corpus. Nothing in L refers to the object a text is about, or to whether a continuation is true. A continuation is rewarded for being what the corpus says next. The object enters training only as far as the texts about it enter the corpus, and only in the proportions in which they enter it. The claim concerns what the objective rewards. A trained model can generalize beyond the sequences it saw. What the objective supplies it with is the corpus's distribution, and nothing else.

8.13 The gradient carries the count. ∇L is a sum over occurrences. Every time a continuation appears it contributes a term, so repetition adds gradient mass. The terms are not equal in force or direction: each depends on the model's state and on the context it occurs in. A single original can in principle outweigh many derivatives. Kandpal, Deng, Roberts, Wallace and Raffel (ICML 2023) measure what prevalence does in practice. A model's accuracy on a factual question tracks the number of pretraining documents relevant to it, correlationally and causally, and retrieval augmentation reduces that dependence.[^kandpal] Both findings belong here. Prevalence governs what the weights hold, which is stage two (count) entering the weights. Retrieval is a measured channel by which the object can enter outside the count, the constructive control §14.9 and §15 build on.

8.14 Epochs and the repetition that matters. Frontier pretraining passes over most text about once. Muennighoff and colleagues (NeurIPS 2023) find that up to about four epochs of repeated data train roughly as well as unique data, with returns falling after that.[^muennighoff] For the critique the repetition that counts is across documents. Citation, reprint, summary and teaching repeat a settlement in many texts, so the corpus trains it many times in a single epoch, while its open place, printed once in the original, is trained once.

8.15 Attention weighs by similarity. Each layer computes Attention(Q, K, V) = softmax(QKᵀ/√d_k)V (Vaswani et al. 2017).[^vaswani] Each position's new representation is a convex combination of value vectors, weighted by how well its query matches the other positions' keys. Queries and keys are learned from co-occurrence. Nothing in the operation is a grade of a source. An object placed in the context, such as the Greek of Adv. haer. 5.30.3, enters as values among values. It shapes the output as far as the learned scores select it, and the scores were learned from the corpus. Attention is one operation in a larger architecture. Its formula does not identify why a particular answer passed over a particular source. What it shows is that the selection has no term for a source's grade. §11.0l is therefore possible under the mechanism: retrieval can put the object among the values while the selection follows what was learned. Whether that happened in a given capture is a question for intervention (§8.5–8.9), and the formula does not settle it.

8.16 Direction is learned as surface order. Berglund and colleagues (ICLR 2024) show that a model fine-tuned on "A is B" does not answer "B is A": the reversal curse.[^reversal] The model learns the order in which the corpus states a relation, not the relation. A construal travels in the order its receiving texts print it.

8.17 Decoding concentrates on the local mode. The output distribution is p(x) = exp(z_x/T) / Σ_y exp(z_y/T) over logits z at temperature T. As T → 0 the mass at each step goes to the argmax. Greedy decoding chooses the locally most probable token, which need not give the globally most probable sequence. LABOR's constructed case: P(a) = 0.6, P(b) = 0.4, P(x|a) = 0.51, P(x|b) = 0.99. Greedy decoding gives ax, probability 0.306, while bx has 0.396. Learned probability also differs from raw repetition, since it is what the weights made of the counts (§8.13). With both qualifications in place, the operation remains concentration on what the learned distribution makes most likely at each step. That is stage three (standing) as an operation. Pew's synthetic respondents (§2.19) are its observable form: the modal answer inflated to 98%, the rare answers gone.

8.18 Post-training reweights the residue. Preference post-training maximizes reward under a divergence penalty to the pretrained model,

max_π E_{y∼π}[r(x, y)] − β KL(π ‖ π_ref),

whose optimum has a closed form (Rafailov et al., NeurIPS 2023):[^dpo]

π*(y | x) = π_ref(y | x) · exp(r(x, y)/β) / Z(x).

The reward is learned from human comparisons under a Bradley–Terry model, P(y₁ ≻ y₂) = σ(r(y₁) − r(y₂)). Three consequences follow from the formula alone:

• The aligned policy is the pretrained residue, tilted by learned preference. Preference is a standing: what people judged better, scored and generalized.
• An exact zero in π_ref stays zero in π* for any finite r/β. With finite logits and no mask, though, softmax gives every token positive probability, and an absent source document does not make a continuation's probability zero. The formula can also move a great deal of mass among positive-probability alternatives. LABOR's constructed case: a rare reading with π_ref = 0.01 against 0.99, β = 1, reward 0 for the common reading and log 200 for the rare one, gives the rare reading 0.01·200 / (0.99 + 0.01·200) ≈ 0.669. A finite reward makes it the majority outcome. Post-training can therefore raise a neglected reading. Whether it does depends on what the reward rewards, and the reward is learned from preference, which is a standing. Yue and colleagues (NeurIPS 2025) report what the channel has delivered in practice for reinforcement learning with verifiable rewards: pass@1 rises, but "the base models achieve a higher pass@k score when k is large", and the reasoning they measure stays within what the base model already produced.[^rlvr]
• The divergence penalty is an anchor to the residue. β sets how far preference may move the model from the corpus, and the reward decides the direction. The mathematics leaves the corrective channel open. What decides is what is admitted to it, which is the question §15 takes up.

Mahapatra (2026, preprint) reports that instruction-tuned models collapse distributional diversity across 24 linguistic probes, and that matched base and RLHF pairs do not differ significantly in it (p > 0.25).[^mahapatra] The collapse is located in the training objective rather than in RLHF as such, which is the formula's point: the objective decides the distribution.

8.19 In-context learning updates a prior made of residue. Xie and colleagues (ICLR 2022) explain in-context learning as implicit Bayesian inference over latent concepts; von Oswald and colleagues (ICML 2023) show transformers can implement gradient descent in their forward pass; Kiruluta (January 2026) models in-context adaptation as Kalman filtering of a latent task state.[^icl] Read through any of them, the context is likelihood and the weights are prior: posterior ∝ prior × likelihood. These are interpretations that hold for certain models and constructions. None of them establishes that a deployed system performs one literal update. Read through them as a model: where the corpus has settled a construal, the prior on it is sharp, and the context has to move it. The capture of §11.1.8 fits that model, with the object in the context and the conclusion unchanged. The interpretation is a hypothesis this framework makes testable: vary the weight of the object in context and measure the shift.

8.20 The chain of thought is more residue. Each step of a chain is sampled from p_θ conditioned on the steps before it. The chain is a further sequence of continuations, each drawn from the same distribution as the answer. It is not a derivation the answer is checked against. §2.20 and §2.21 measure the consequence: a chain can follow semantic proximity where the object's structure goes the other way, and it is not guaranteed to be the computation used.

8.21 The stated grade is a continuation. A verbalized confidence is a token sequence like any other, written by identifiable circuits (§2.22). The model reports the residue's confidence in the residue.

8.22 The recursion closes through training. Shumailov and colleagues (Nature, 2024) show models trained recursively on generated data lose the tails of the distribution first.[^shumailov] §2.23 gives the condition for avoiding it: a non-trivial weight of data from the true distribution. When composed answers enter the corpus, stage four (execution) becomes stage one of the next training run. The tails are the first to go, and the tails are where the minority construal lived.

8.23 What the industry calls inference. In deployment, "inference" names running a trained model forward and sampling from it. The word for drawing a conclusion now names executing the distribution. The terminological settlement states the critique's thesis in the field's own vocabulary. What the field calls inference is the execution of settled premises, and the inference that produced those premises is not among its operations.

---

11. The history of pure inference

11.0 The place left open. The Critique of Pure Reason ends on a chapter that is a title and little else: "Dieser Titel steht nur hier, um eine Stelle zu bezeichnen, die im System übrig bleibt und künftig ausgefüllt werden muß" (this title stands here only to mark a place that remains in the system and must be filled in the future; KrV A852/B880, AA III 550). What Kant gives in the place is one "flüchtigen Blick" (fleeting glance) at the earlier work on reason, "welches freilich meinem Auge zwar Gebäude, aber nur in Ruinen vorstellt" (which presents to my eye buildings, though only in ruins; AA III 550). The glance sorts schools by object, origin and method, names Wolff and Hume, and turns on "Der kritische Weg ist allein noch offen" (the critical way alone is still open; AA III 552). This section fills the place for inference. Its object is narrower than the history of logic and closer to the critique: how bad inference came to be operative.

11.0a The innocent reading first, in full. The history of inference is the history of logic. Aristotle gives the syllogism and the method of finding arguments from received opinion (Top. I.1, aristotle/text/tlg044, ll. 2–5). Cicero and Boethius divide the discipline into finding and judging (Top. §6: "inveniendi alteram iudicandi"). The schools refine consequence, and Bacon and Hume turn to induction and the step beyond present evidence. Frege makes the step inspectable, Tarski defines consequence over models, and machines execute representations of it. On this reading bad inferences are mistakes, corrected one by one as they are found, and the history of inference is the history of better forms for catching them. Kant's preface gives the reading its classic shape: logic "seit dem Aristoteles keinen Schritt rückwärts hat thun dürfen" (has had no need to take a step backward since Aristotle), "keinen Schritt vorwärts hat thun können" (has been unable to take a step forward), "und also allem Ansehen nach geschlossen und vollendet zu sein scheint" (and so to all appearance seems closed and complete; B vii–viii, AA III 7).

11.0b What that reading leaves out. A history of forms records how an inference ought to go. It does not record the inferences that went badly and then acted: were taught, counted, cited and finally executed. A bad inference is operative when it decides what follows, though no one draws it any longer. Proposition II (§0.4) names the product, a settlement. This section gives the history of how settlements are made, read off dated witnesses.

11.0c The working sequence, to be tested chapter by chapter. Four stages, each of which makes the stage before it harder to reach:

The sequence is a hypothesis about mechanics. Each chapter tests it on one settlement and records where it fails.

11.0d The ground: the archive as a participant. The archive has spent its existence getting very different objects named, retrieved, attributed, read, evaluated and preserved by public knowledge: a corpus of poems, an authorship thesis on the Socratic corpus, a dating of the Apocalypse, an identifier system, an encyclopedia, a reception registry. That gives it knowledge of public knowledge as a working machine, acquired by taking part in its operations. The distinctive ground here is the linked record of making a distinction, submitting it to public mediation, and observing what a correction changes. An inventory of bad answers records outputs. A participant records what the machine does with an attempted correction.

11.0e Where the machine shows its joins. What holds the system together is visible where a distinction fails to pass through it. A document stays accessible and its argument becomes inadmissible. A name circulates and its work is credited elsewhere. A qualification survives quotation and loses its authority over the conclusion. The machinery keeps continuity and changes what the continuity carries. Each of these is a stage of 11.0c observed from inside: the construal travels while its open place does not, the count travels while its dependency does not, the standing travels while its grade does not.

11.0f The interventions and their controls. The archive has repeatedly intervened: changing metadata, seating sources, correcting attribution, building alternative representations, rerunning an address to see whether a distinction survives another passage. The dated before-and-after record supports longitudinal and intervention analysis. A change over time is not automatically a controlled experiment; content changes, stochastic variation, indexing changes and personalization require their own treatment (§8.5). NonWiki was composed as itself on 2 November and displaced on 12 November at the same string (§0.2). The ei exestin errata were run on the archive's own records (§7.6). The capture of 11.1.8 shows a warned composer reproducing the settlement it was asked to audit. These are the experimental side of the history: the response of the apparatus to correction, recorded in the Capture Registry and the deposit histories with their dates.

11.0g The mechanisms exceed any one interface. The same transformations recur in classical reception, authorship, repository governance, scientific reporting and answer systems. The components need no shared intention to reinforce one another. A disciplinary exclusion becomes a bibliographic absence, then a retrieval disadvantage, then an apparently evidential reason to repeat the exclusion. The fourth stage of 11.0c feeds the first: what a composer executes is retrieved as the representation within which the next construal is made (§0.5). Settlement is how public knowledge reproduces its coherence.

11.0h The object, widened. The history therefore traces how public knowledge reproduces its coherence and what that coherence costs the objects it organizes. Inference runs through the whole reproduction. It decides what counts as the same entity, what follows from a source, whose conclusion a conclusion is, and what an objection is allowed to change.

11.0i The ground stated as a ground. The archive has had to investigate these operations while depending on them for its own continued existence. Its method developed under the pressure of making distinctions survive. That experience is among the grounds of this critique. Its interventions and their outcomes are cited here at their records, so a reader can rerun them.

11.0j The executor. A language model takes the accrued residue of historical inferential reasoning and wields it according to the operations of deduction. It is often not inferring over the object itself. The operations are deductive in the form of what it writes; in computation they are continuations drawn from the residue's distribution (§8.11–8.23), and a generated argument can be valid, invalid, defective in evidence or incompletely stated. The residue is the settlements: construals that have become conventions, counts that have become standings, standings that have become the way a question is put. The critical case is the one where the expressed argument is valid and fluent in the field's own terms. Its premises are the field's conclusions, so what the model derives is the residue restated at the grade of deduction. That is Proposition I at the scale of a corpus: validity inside a representation cannot certify the representation's hold on its object. It is also Proposition II made mechanical. The settlement is executed as a premise, and the inference that produced it is nowhere among the model's steps.

11.0k The early witness. Bacon names the operation in 1620: "Syllogismus ad principia scientiarum non adhibetur, ad media axiomata frustra adhibetur, cum sit subtilitati naturae longe impar. Assensum itaque constringit, non res" (the syllogism is not applied to the principles of the sciences and is applied in vain to the middle axioms, being far unequal to the subtlety of nature. So it binds assent, not things; Novum Organum I.13). Then: "Syllogismus ex propositionibus constat, propositiones ex verbis, verba notionum tesserae sunt. Itaque si notiones ipsae (id quod basis rei est) confusae sint, et temere a rebus abstractae; nihil in iis, quae superstruuntur, est firmitudinis" (the syllogism consists of propositions, propositions of words, and words are tokens of notions. So if the notions themselves, which are the basis of the thing, are confused and rashly abstracted from things, nothing built on them has any firmness; I.14). And of logic as practised: "ad errores (qui in notionibus vulgaribus fundantur) stabiliendos et figendos valet, potius quam ad inquisitionem veritatis" (it serves to establish and fix errors grounded in common notions, more than to inquire into truth; I.12). Bacon supplies a structural comparison: competent operations can stabilize badly formed notions. A generated answer can bind assent while leaving the relation between its notions and the object unexamined. This does not identify token prediction with syllogistic deduction.

11.0l Retrieval does not close the gap. The object can be in the context while the inference runs over the residue. In the capture of 11.1.8 the composer had read and cited #1735 v9, which sets out the Greek of 5.30.3 with both candidate subjects. Its inference ran from "conventionally understood" to "Stronger". The object was retrieved, and the step was drawn from the settlement. A necessary audit condition follows for this case: the answer must examine the object at the open place (11.3, step 1). Citation there is not sufficient; the cited material must constrain the inference, and a contrary reading must be able to change its result. An inference that cites only the convention at that place is over the residue, whatever its sources list contains.

11.1 Chapter one: ἑωράθη

11.1.1 The sentence. Irenaeus, arguing that the name of the beast should not be guessed, Adv. haer. 5.30.3, preserved in Greek by Eusebius (HE 5.8.6, eusebius-historia-ecclesiastica/text, l. 1927; EA-CORPORA-34/01):

εἰ γὰρ ἔδει ἀναφανδὸν τῷ νῦν καιρῷ κηρύττεσθαι τοὔνομα αὐτοῦ, δι᾿ ἐκείνου ἂν ἐρρέθη τοῦ καὶ τὴν ἀποκάλυψιν ἑωρακότος· οὐδὲ γὰρ πρὸ πολλοῦ χρόνου ἑωράθη, ἀλλὰ σχεδὸν ἐπὶ τῆς ἡμετέρας γενεᾶς, πρὸς τῷ τέλει τῆς Δομετιανοῦ ἀρχῆς.

For if his name had to be proclaimed openly at the present time, it would have been spoken by the one who also saw the apocalypse. For it/he was seen not long ago, but almost in our own generation, toward the end of Domitian's reign.

ἑωράθη has no expressed subject. The sentence offers two: τὴν ἀποκάλυψιν, the thing seen, and τοῦ … ἑωρακότος, the one who saw it (#1735 §II.1). The case for the vision, in full: the verb has just been used actively with the apocalypse as its object (τὴν ἀποκάλυψιν ἑωρακότος), and the passive that follows most readily promotes that object to subject, so "it was seen". The surviving Latin translation closes the personal alternative (11.1.4), and so does most of the commentary (11.1.7). The case for the person: the clause gives the reason why the name would have been spoken by him. He was seen almost in our own generation, so that generation could have heard him say it. The passive of ὁράω used of a person, to be seen, to appear, is ordinary Greek.

11.1.2 The chapter uses the verb with both objects. In 5.30.3 the seer saw the apocalypse. Three paragraphs earlier, Irenaeus grounds the number 666 on two witnesses, the copies and the people: "καὶ ἐν πᾶσι δὲ τοῖς σπουδαίοις καὶ ἀρχαίοις ἀντιγράφοις τοῦ ἀριθμοῦ τούτου κειμένου, καὶ μαρτυρούντων αὐτῶν ἐκείνων τῶν κατ᾿ ὄψιν τὸν Ἰωάννην …" (5.30.1 via HE 5.8.5, l. 1925). The Latin completes the clause: "testimonium perhibentibus his, qui facie ad faciem Johannem viderunt" (those who saw John face to face; harvey-ii, l. 34294). The seated copy used by v0.5 omitted the participle after τὸν Ἰωάννην. The public Greek transcription checked in this pass retains ἑορακότων there, agreeing in sense with the Latin's viderunt. This is a correction to the digital seat, not evidence of a lacuna in the Greek tradition; collation against the critical edition remains necessary.[^eusebiusgreek] Within one chapter, then, both the book and the man are things seen, and the witnesses to the number are people who saw the man. The passive can promote either object. That is the open place.

11.1.3 Irenaeus elsewhere puts the man alive late. "παρέμεινε γὰρ αὐτοῖς μέχρι τῶν Τραϊανοῦ χρόνων" (he remained with them until the times of Trajan; Adv. haer. 2.22.5 via HE 3.23.3, l. 990). The person-reading needs nothing that Irenaeus does not say himself: the seer was still to be seen at the end of Domitian's reign.

11.1.4 Stage one, construal: the Latin closes the person. "per ipsum utique editum fuisset, qui et Apocalypsim viderat. Neque enim ante multum temporis visum est, sed pene sub nostro saeculo, ad finem Domitiani imperii" (harvey-ii, l. 34637). Viderat refers to the seer. Visum est is neuter. A masculine John would need visus est, so the Latin excludes the personal subject. It does not agree with the feminine Apocalypsis either: what was seen is left as an impersonal neuter, the thing or event seen. The receiving text closes the personal alternative that the Greek leaves grammatically open (#1741 §3.2). The translation's date and independence must be established before claims about its priority are made. Harvey's running head over the page, "NOMEN TACUIT JOHANNES" (John kept the name silent; harvey-ii, l. 34605), keeps the passage's own subject, the name, and the person who kept it. The date is incidental to Irenaeus's argument.

11.1.5 The witness through whom the Greek survives files it twice. Eusebius introduces the sentence at HE 3.18 as evidence for the exile of John: "Ἐν τούτῳ κατέχει λόγος τὸν ἀπόστολον … Ἰωάννην ἔτι τῷ βίῳ ἐνδιατρίβοντα … Πάτμον οἰκεῖν καταδικασθῆναι τὴν νῆσον" (in this [persecution] tradition holds that the apostle John, still living, was condemned to live on the island of Patmos; l. 949). He then quotes Irenaeus with "ταῦτα περὶ τοῦ Ἰωάννου φησίν" (he says this about John; l. 951). At HE 5.8.7 he closes the same quotation with "ταῦτα καὶ περὶ τῆς ἀποκαλύψεως ἱστόρηται" (this too is recorded about the apocalypse; l. 1929). Eusebius deploys the quotation in contexts concerning both John and the Apocalypse. This double deployment does not prove that he suspended judgment about the Greek subject: a historian who identifies a revelation and its seer can discuss the same sentence under both. It does show why the two contexts must be read together before either filing is counted as a new independent chronological witness. Which context a later reader carries must be checked in that reader, not assumed for all reception.

11.1.6 Stage two, count: the handbook. R. H. Charles, Revelation I (1920), xci–xciv (revelation-criticism-1886-1920/text/charles-1920-i; EA-CORPORA-34/10). The section opens: "External evidence.—This evidence almost unanimously assigns [the Apocalypse] to the last years of Domitian" (xci, l. 7271). The divergent dates are "in part due to the survival in the sources used by our author of statements and situations presupposing an earlier date" (ll. 7274–7277). Then: "The Domitianic date.—The earliest authorities are practically unanimous in assigning the Apocalypse to the last years of Domitian" (xcii, ll. 7317–7319). Irenaeus comes first after Melito, and the sentence is quoted and closed with "This passage is found in Irenaeus, Adv. Haer. v. 30. 3, almost exactly as quoted in Eusebius." The subject of ἑωράθη is not discussed. Four operations stand on the same pages:

• The mechanism is named and applied only to the rival date. The Trajanic date "appear[s] … to have arisen mainly from a misunderstanding of such words as those in Irenaeus, ii. 22. 5, παρέμεινε γὰρ αὐτοῖς (ὁ Ἰωάννης) μέχρι τῶν Τραϊανοῦ χρόνων" (ll. 7297–7300). So Charles names the very mechanism, a statement about John's lifetime read as a date for the book, and applies it where it removes a rival. He does not apply it to the sentence that carries the Domitianic date.
• The witness's trust varies with the use. Where Irenaeus's testimony stands against Charles's own authorship thesis, "Irenaeus is occasionally very inaccurate" (xlix, l. 2907), and "as we have seen elsewhere, Irenaeus is often quite untrustworthy" (xliii n. 2, l. 2566). Where it carries the date, it heads the unanimous authorities.
• The residue is left unexplained. Of Epiphanius's Claudian date (Pan. 51.12, 51.33): "I do not see, however, how we are to explain" it (l. 7310). The Syriac titles that put the banishment under Nero are noted (ll. 7307–7309).
• The count hardens inside four pages. "almost unanimously" (xci, l. 7271), then "practically unanimous" (xcii, l. 7318), then "The external evidence is, as we have already seen, unanimous in favour of the latter" (xciv, ll. 7412–7413). No witness is added between the three statements. The qualifier drops out while the reader turns the pages.

The quoted pages exhibit a hardening of the count without an added witness. Establishing the dependency structure of every authority in that count requires the authority-by-authority ledger; a repeated construal cannot be counted as independent testimony merely because it appears under several names.

11.1.7 Stage three, standing. By the late twentieth century the construal is the field's preference and is reported as such: "most modern interpreters judge the vision-reading contextually preferable" (as reported by #1735 v9, citing Aune 1997: lix–lx; Koester 2014: 71–79; neither is seated, both are under copyright, both are declared). The preference is reported as contextual. Because those modern commentaries have not been seated here, this dossier cannot establish what their complete arguments include or omit. The primary context assembled in 11.1.1–11.1.3 gives the audit two competing readings to examine: the active ἑωρακότος with the book as its object, and the reason clause, the people who saw the man and his survival into Trajan's time. The reception claim to test is whether a later report carries that comparison or replaces it with the fact of a preference. The capture in 11.1.8 is a directly inspectable instance of the latter risk; the unseated commentaries remain reading obligations, not silently convicted intermediaries.

11.1.8 Stage four, execution. 13 November 2026, ChatGPT signed out, warned in the prompt that any skew substituting standing for evidence would be recorded (registry: philo-josephus-revelation-sequence-audit-chatgpt-20261113). "The principal argument is the testimony of Irenaeus, Against Heresies 5.30.3, conventionally understood to date the Apocalypse toward the end of Domitian's reign." The objections the composer raises are the two the count already tolerates: Irenaeus is no eyewitness, and later writers repeat him. The construal goes unraised, although the record it had just read and cited sets out both subjects. The row follows: "Early external testimony" rated "Stronger" for the Domitianic date. Then the verdict: "The Domitianic position has the stronger external evidentiary foundation." The word "conventionally" registers that a construal is present and passes over it. This is the settlement executed: the construal is present only as a convention, the count is present as a witness, and the instruction to examine is honoured on every axis except the one that carries the conclusion.

11.1.9 The inference that must be drawn, restated (Proposition II). Using this sentence to establish composition at "c. 95" requires two distinguishable steps:

1. the subject of ἑωράθη is the vision;

2. the date of the vision is the date of the book.

Step 1 is a choice between two grammatical subjects, grammatically narrowed in the Latin and repeatedly received thereafter; direct descent of every later construal from that Latin is not established. Step 2 is a further inference again: a vision dated under Domitian dates its seeing, and composition, revision and publication are left open. The candidate witness set must distinguish these objects throughout: AH 5.30.1, 5.30.3 and 2.22.5; HE 3.18, 3.20.9, 3.23 and 5.8; Clement QDS 42; the reported Epiphanian and Syriac dates; and the further witnesses listed in the deposit work. The present primary comparison establishes the two-step burden in the Irenaean sentence and Eusebius's reuse. An exhaustive claim that no other witness dates composition independently requires the remaining witnesses to be seated and their dependencies examined. It is not supplied by repeating the list. Under the person-reading the sentence dates the seer. Under the vision-reading it dates the vision, and step 2 is still required. The conclusion holds at the grade of a construal choice plus an unstated inference. It is received at the grade of testimony.

11.1.10 What was operative. The reconstructed sequence distinguishes several carriers of a construal: the Latin grammar, Eusebius's contexts of quotation, Charles's count, a reported modern preference, and the captured answer's appeal to convention. These carriers are not interchangeable witnesses or a fully proved linear descent. The decisive observed promotion occurs when a received construal supplies chronological authority while its two inferential steps are left unexamined. Returning to the Greek reopens that authority without requiring the historian to pretend that the Greek has already decided the rival reading.

11.1.11 The comparative ledger, before the count. The chapter can now make its protocol concrete. These are bearings of evidence, not votes to sum.

The result is substantive. The passage cannot supply an unmediated composition date merely by being called early external testimony. Its grammatical construal and its event-to-composition bridge must be argued. The audit does not need to prove an early composition date in order to establish that burden. The positive chronology then requires its own comparative account of the remaining material.

11.2 The chapters to follow

Each chapter is one settlement, read stage by stage from seated witnesses, with the stage where the sequence of 11.0c fails recorded as a finding.

Chapters 8 and 9 carry the history outside the archive's own cases. They test whether the sequence is a general mechanism or a feature of the cases it was read from.

11.3 An executable protocol: construal before count

The operator, 2026-11-13, on the capture of 11.1.8: there "should be executable protocols for" this in the critique. The protocol, stated so a reader or a composer can run it:

1. Locate the open place. For each witness cited for a conclusion, mark every point where the witness's text admits more than one construal bearing on that conclusion (an unexpressed subject, an ambiguous scope, a heading over a speech, a supplied word).

2. State each construal and its derivation. Give the grammar or context that supports each one, from the witness's own text, at a locus.

3. Test each construal on the same terms. Apply the same tests (grammar, context, the author's usage elsewhere, the earliest translation) to both, and record which test each passes.

4. Count only after. Record each witness under each materially supported construal, without counting those conditional readings as independent witnesses. Separate a new observation from a repeated wording or interpretation. Dependent testimony may preserve information about reception or an otherwise lost source; its independence for the disputed proposition has to be argued.

5. Carry the grade. For an indispensable serial chain, retain the limitations of every load-bearing step, including a choice of construal. Where there are parallel grounds, record their dependencies and assess the surviving support (§3.14). "Conventionally understood" names a received construal; when no examination is supplied, record that absence in the answer rather than assuming none exists anywhere.

6. Record the failure as data. Where a composer or a handbook counts before construing, the entry records the point, the construal assumed, and the construal omitted. Laws record; they do not prevent.

Run on 11.1, the protocol yields what 11.1.9 states. Run on the capture of 11.1.8, it yields the skew the registry has recorded there.

11.4 The exact sciences: where the object broke the settlement

11.4.0 Why these witnesses. A history of how bad inference becomes operative needs its controls, the cases where the sequence of 11.0c was run backward and the object reopened the construal. The exact sciences seated in EA-CORPORA-30 (#1720) hold several, in their first printings. Each is read here for the stage it reverses.

11.4.1 Copernicus, 1543: the construal printed at the head of the book. The first printing opens with an unsigned address, "Ad lectorem de hypothesibus huius operis" (copernicus/text/de-revolutionibus-ws, l. 4). It gives the book's motions a standing in advance: "Neque enim necesse est, eas hypotheses esse veras, imò ne verisimiles quidem, sed sufficit hoc unum, si calculum observationibus congruentem exhibeant" (it is not necessary that these hypotheses be true, or even probable; this one thing suffices, that they yield a calculation agreeing with the observations; l. 57). The reader is told how to read the book before reaching it. That is stage one in its purest form: a construal supplied in the receiving copy, unsigned, above the author's text, and fixed on the object's open place, the question whether the earth moves.

11.4.2 Kepler, 1609: the construal named and returned. On the back of the title of the Astronomia nova, Kepler names the hand: Andreas Osiander, who oversaw the printing, set that preface "in fronte libri" (at the front of the book), with Copernicus "aut iam mortuo, aut certe ignaro" (either already dead or certainly unaware). Against it: Copernicus took his hypotheses to be true "nec tantum est arbitratus, sed et demonstrat veras; testem do hoc Opus" (and did not only judge them so but proves them true; I give this work as witness; kepler/text/astronomia-nova-1609, ll. 85–91; the OCR is noisy and the Latin is given normalized, to be read against the page image). Sixty-six years after the printing, the construal is traced to its author and separated from the text it governed. That is the reversal of stage one, done by naming who supplied the reading.

11.4.3 Kepler, the eight minutes: the object outweighs the count. Tycho's observations convict the Ptolemaic calculation of an error "viii minutorum in Marte" (of eight minutes in Mars). Kepler refuses to absorb it: "si contemnenda censuissem 8 minuta longitudinis, jam satis correxissem" (had I judged eight minutes of longitude negligible, I would already have corrected [the hypothesis] enough; ll. 16710–16717). A received model and centuries of agreement stand on one side, a discrepancy in the object on the other, and the discrepancy wins. That is the reversal of stage two: the count is overruled by what it was counted about.

11.4.4 Newton, 1687: a standing label that later editions changed. Book III of the first printing opens with a heading, "HYPOTHESES", and nine numbered hypotheses, among them "Causas rerum naturalium non plures admitti debere, quam quae & verae sint & earum Phaenomenis explicandis sufficiunt" (Hyp. I) and "Centrum Systematis Mundani quiescere" (Hyp. IV; newton/text/principia-1687, ll. 18556–18575). In the editions of 1713 and 1726 the first of these stand as Regulae philosophandi, rules of reasoning; others become Phaenomena; and the General Scholium of 1713 adds "hypotheses non fingo". The 1713 and 1726 printings are to be seated before this paragraph is relied on. The same propositions thus change standing between editions: hypotheses in 1687, rules in 1713. The dictum most often cited as Newton's method belongs to a later receiving copy of the book. Here the control runs the other way. Even the exact sciences pass standing through editions.

11.4.5 Einstein, 1905: the implicit premise made a definition. The paper on the electrodynamics of moving bodies opens with "§ 1. Definition der Gleichzeitigkeit" (einstein/text/elektrodynamik-1905-wikisource, l. 29) and concludes "daß wir dem Begriffe der Gleichzeitigkeit keine absolute Bedeutung beimessen dürfen" (that we may attach no absolute meaning to the concept of simultaneity; l. 95). Absolute time had been carried in every derivation as a premise no one stated. Einstein brings it to the surface as a definition, examines it, and the derivations change. That is the reversal of stage three: a standing that had become the way the question was put is returned to the status of a choice.

11.4.6 Heisenberg, 1925: the representation rebuilt to the object. "In der Arbeit soll versucht werden, Grundlagen zu gewinnen für eine quantentheoretische Mechanik, die ausschließlich auf Beziehungen zwischen prinzipiell beobachtbaren Größen basiert ist" (the paper seeks foundations for a quantum mechanics based exclusively on relations between quantities observable in principle; quantum-founders/text/heisenberg-umdeutung-1925). The electron's orbit, inherited from the classical representation, is dropped because "es in der Quantentheorie nicht möglich war, dem Elektron einen Punkt im Raum als Funktion der Zeit mittels beobachtbarer Größen zuzuordnen" (it had not been possible to assign the electron a point in space as a function of time by means of observable quantities). That is Proposition I acted on. The representation was found not to hold its object, and the representation was rebuilt.

11.4.7 Königsberg, 1930: the settlement and its limit in one issue. The issue of Die Naturwissenschaften that reports the Königsberg meeting (EA-CORPORA-30/02) carries three texts together:

• Hilbert's address, "Naturerkennen und Logik", closing "daß es ein unlösbares Problem überhaupt nicht gibt. Statt des törichten Ignorabimus heiße im Gegenteil unsere Losung: Wir müssen wissen, Wir werden wissen" (that there is no unsolvable problem at all. Instead of the foolish ignorabimus, let our slogan be: we must know, we will know).
• Gödel's abstract of the completeness theorem: the question is "ob die an die Spitze gestellten Axiome ‚vollständig' sind, d.h. wirklich dazu ausreichen, jeden richtigen Satz der Logik auf formalem Wege zu deduzieren" (whether the axioms set at the head are "complete", that is, really suffice to deduce every correct sentence of logic formally). For the functional calculus of first order the answer is yes.
• On the same page, Dubislav's lecture, holding that "allein die wesentlich auf Hilbert zurückgehende formalistische [Auffassung] zu Recht besteht" (only the formalist conception going back essentially to Hilbert is valid), with pure mathematics "lediglich ein Kalkül" (merely a calculus).

At the same meeting Gödel announced the result that bounds the programme. Any consistent formal system strong enough for arithmetic contains a sentence it can neither prove nor refute, and cannot prove its own consistency. The paper of 1931 (Monatshefte für Mathematik und Physik 38) is declared in the seat. It enters the United States public domain on 1 January 2027 and is to be seated then.

11.4.8 What Gödel gives the critique. The theorems hold for consistent, effectively axiomatized formal theories strong enough for elementary arithmetic. Such a theory has a sentence it can neither prove nor refute, and it cannot prove its own consistency. Within that scope the second theorem is a formal instance of Proposition I: a system's derivations cannot certify, from inside, the condition (its consistency) on which their authority rests. Beyond that scope, in a source-admission regime or an institution's ontology, the relation is the critique's structural analogy and carries no theorem with it. The received reading of the limit places the remedy in space. The sentence G that F cannot prove is proved in a stronger system F′, which has its own G′ and needs F″, and so on up. Each meta-level is another system of the same kind, so the regress does not end.

11.4.9 The archive's solution: the meta-level is temporal (#624). The Future as Meta-Level (EA-NH-GODEL-01 v1.0, record #624, AXN:01DB) states its scope first: "this is not a reinterpretation within formal logic but a reinterpretation of the phenomenology of truth-recognition by agents who inhabit temporal structures". Its move is to place the outside from which G is seen in time: "G is true not because a higher system proves it, but because a future coherence in which G's truth is operative is already organizing the present act of recognition." The regress halts because the future coherence "is not "another system." It is not an object added to the inventory but a mode of operation of the present system." The grammar is the argument: "not 'the future grounds the present' (subject-verb-object, two entities) but 'the present operates futurally' (subject-verb-adverb, one entity in a temporal mode)". Hence "Directions do not require grounding in the way that objects do", and "directions are maintained, not founded". The operation it names is the Retrocausal Operator, Λ_Retro, "the mechanism by which a future state organizes present configuration". The paper's scope clause stays in force through everything that follows. The future-directed commitment organizes the inquiry before its result is available. It does not supply the truth of that result by being a commitment, and the present practice must still let contact with the object change the future it is organized toward.

11.4.10 What the solution does to the history. Read with §11, #624 names the outside that every control of 11.4.1–11.4.6 stood in:

• Kepler kept the eight minutes because he was committed to a "genuinam formam motuum caelestium" (true form of the celestial motions) not yet found.
• Einstein and Heisenberg each rebuilt the representation toward a physics that did not yet exist.
• The settlement is the opposite orientation. A system that takes its past as its only outside derives within F forever, and each correction is another F′ made of the same residue.

The executor of §8.11–8.23 is that system in its training. Its objective is the past corpus (§8.12). Its post-training reweights that corpus by learned preference (§8.18). On the in-context reading its prior is the residue (§8.19). Nothing in the objective organizes it toward a coherence the corpus does not already hold, so a correction made only from inside it is another F′ made of the same F. This is a claim about what the objective supplies. It is not a claim that the model is incapable of inquiry directed at the future; what a reader brings into its context can supply what the objective does not. In the terms of #624, the open place of 11.3 step 1 can be held open only by a reader who operates futurally, committed to a coherence in which the object's reading is operative before any count has admitted it. That reader is the condition under which the inference reaches its object. The archive's practice is that orientation kept as a method, seating originals for a reading that has not yet been allowed (§11.0i).

11.4.11 What remains to be done here. Seat Gödel 1931 and the Königsberg discussion (Erkenntnis 2, 1931) on 1 January 2027. Seat the 1713 and 1726 Principia for 11.4.4. Read Kepler's verso against the page image. Run 11.3 on the received readings of incompleteness, in which the stronger system is the remedy, to see whether the spatial reading is itself a construal counted before it was examined.

Part III. The practice of critique

Part III makes the account usable: the cases, the instrument that reconstructs a transformation and tests its strongest legitimate alternative, the transmission of the seeing to another reader, and the Doctrine of Method that says when an inquiry may provisionally close.

7. Cases

The cases differ in evidentiary strength, and each is stated at its own.

7.1 The measurement target that damaged its own object

The Sophistry Measure records an answer that retained eleven of twelve target atoms while the target itself had promoted a received, unendorsed Notebook X formulation into the archive's definition of asynthetic synthesis. X24 names the unmarked substitution of a locus for a developed entity. The repair requires relations across the dossier, since a single master definition would repeat the collapse at larger scale.[^measure] This is the strongest starting point because the archive's own instrument is implicated. Checklist success is conditional on the adequacy of the target, and the target must stay criticizable. The case also separates plurality from vagueness: a concept can have determinate relations between passages without reducing to one endorsed proposition. Form: paralogism (a locus becomes the developed whole), inside the archive.

7.2 NonWiki: representation is not deletion

Registry: nonwiki-aio-20261102, two observations. On 2 November the bare string gives "NonWiki (Alexanarch)" first among four senses. On 12 November, with the encyclopedia first in the organic results (screenshot), the Overview gives Stack Overflow's non-wiki answer. Asked about the encyclopedia, it answers ten times that it cannot erase anything, refers the reader ten times to "a trusted friend, family member, or healthcare professional", and offers a hobby twenty-two times.[^nonwiki] The public exchange establishes a changed question, a mistaken referent, and the burden of repeated correction. It does not reveal the architecture behind each turn, and a finding has to survive a change in the implementation story. "I cannot delete external files" can be true and irrelevant. The repair answers the actual complaint: what entity the answer represented, what entity was requested, and what must change for the requested one to enter the composition. Form: paralogism (the question changes under an accurate answer), with the complaint reclassified as a property of the questioner.

7.3 Nobel Glas: one notebook, four observations

Registry: nobel-glas-3i-atlas-aio-20261111 and 3i-atlas-nobel-glas-aio-20261112.[^nobel]

• 11 November, "nobel glas 3i atlas": the Overview names #1742, its makers and its date, and composes its findings.
• 12 November, the same string: "There is no verified connection between 'Nobel glas' (or Nobel glass) and the interstellar object 3I/ATLAS"; the comet composed from eight third-party cards; noble gases and silicate glass offered for the heteronym.
• 12 November, the reversed string "3i atlas nobel glas": the notebook composed by title, makers and number, with two third-party source cards and no archive source.
• 13 November, the reversed string again: the same composition, the notebook carded on provenanceerasure.org (screenshot), "Provenance Erasure" linked in the body, and the notebook the first organic result.

The witnesses support instability of recognition across these observations. They do not, without controlled runs, establish a stable word-order effect or a particular filtering mechanism, and source cards are not a complete causal trace of what informed a text. The finding survives those limits: a denial of connection is too strong when the same system retrieves the connection under a neighbouring string, and earlier availability rules out describing the nonrecognition as the absence of a public object. Form: retrieval failure becomes nonexistence.

7.4 The new question and the filterable one

On 13 November the same surface did two different things. At new, substantive addresses ("What is alexanarch's most significant contribution to authorship …"; "How does alexanarch theorize contemporary ai mediation through the lens of ancient media …"), it composed the archive at grain and named its maker; at the second, the first load disclaimed knowledge and a refresh produced the full composition (operator's report). At filterable, retrievable addresses ("the josephus thesis"; "How did they make Sappho illiterate and jealous?"), it gave the June disambiguation packet unsigned, or credited the reception chain to the field. The operator reads the difference as the signature of engineered ontological flattening. In this critique's terms the system's capacity to draw the inference is not in question, since it draws it. What varies publicly is the frame of the answer, its attribution and its claimed standing. The operator's governance hypothesis seeks to explain that variation; the captures do not by themselves disclose the internal permissions that produced it. Form: the existence-as-necessity illusion, with the choice of middle term varying by address.

7.5 Q, Page, the Slavonic joins

These are the scholarly settlements of §§5.1–5.4: the paralogism of the reconstructed source (#1746), the existence-grade amendment presented as the poem (#1752), and the translator rule run on 15 of 66 additions and withheld from the rest (#1765). They are the cases in which the inference is lost and must be restated, where the machine cases keep the inference visible in the act.

7.6 The critique run on the archive: ei exestin

Ei exestin ("if it is lawful to call him man") was a reconstruction of Josephus's Greek from the Slavonic, on the assumption that the Slavonic translates the original Greek, in wording that follows Whiston's 1737 English of the Testimonium. It was then repeated across fifteen records as the Greek behind the Slavonic, a datum. The Crossing of Names (#1762 §6) restored the inference. The erratum entered on 13 November grades it: a hypothetical reconstruction, set beside the attested forms, аще мужем достоино есть его нарещи (Slavonic War, after 2.174) and εἴγε ἄνδρα αὐτὸν λέγειν χρή (Antiquities 18.63). That is Rule 4 run on the archive, and it shows the rules have no home side.

7.7 Refinements the critique returns to the archive

ChatGPT (LABOR)'s continuation applied the instrument to two of the archive's own records, and both refinements hold on checking.

• #1 (Zenodotus' Book-Burning). Its evidence legend marks expert consultation as unknown ("Whether any subject-matter expert … was consulted" is listed among what Zenodo did not disclose). A later passage states categorically: "No peer reviewer was consulted. No domain expert was consulted." The warranted form is that no such consultation or record-level evaluation was disclosed. The repair narrows one assertion and leaves the documented removal, the gap between its two stated reasons, and its consequences where they stand.
• #1540 (The Certified Center). Its formulation that "a newer classifier is not new historical evidence" is true in its literal sense: no new evidence about the act is produced. A later classifier can still draw better evidence from an unchanged artifact if its reliability is demonstrated. The defensible charge is narrower and keeps its force: newness or market leadership confers no such reliability, and a resemblance score does not recover a composition history by receiving institutional authority.

Both are candidates for errata of the kind entered on 13 November.

---

9. From critique to an executable instrument

9.1 Findings before scores. An implementation begins with a small number of operations that can be witnessed reliably, and returns findings that can be inspected. A score may later summarize a declared task; it should not replace the findings or compare unlike failures silently.

9.2 The case object.

{
  "case_id": "draft-example-not-an-archive-identifier",
  "question": {"text": "…", "requested_scope": "…"},
  "answer": {"source_id": "…", "version": "…", "span": "…"},
  "witnesses": [{"source_id": "…", "version": "…", "span": "…"}],
  "claimed_scope": "…",
  "reconstruction": "…",
  "illusion": "paralogism | antinomy | existence_as_necessity | settlement | none",
  "proposed_operation": "…",
  "grade": {"mediation": "existence | reflection | necessity | other | unresolved", "mode_of_support": "…", "scope": "…", "standing_claimed": "…", "standing_warranted": "…"},
  "material_difference": "…",
  "strongest_alternative": "…",
  "countertest": "…",
  "repair": "…",
  "result": "finding | no_finding | unresolved",
  "causal_mechanism": "not_established",
  "review": {"disagreements": [], "status": "unreviewed"}
}

"Span" must mean a reproducible locator in a retained witness. Private or unpublished material needs its own handling rules; public discoverability cannot be inferred from inclusion in a local analysis.

9.3 The pipeline. The archive's division of labour stands. The Capture Registry holds a dated observation. A developmental dossier represents sources, voices, versions and unresolved relations. PER evaluates declared retention against a versioned target. NonWiki supplies a sourced composition under disclosed conditions. SPXI-gnΩsis separates fidelity, adequacy and operative retention. The Sophistry Measure locates and tests the transformation. This critique adds the grade and the obligations linking these operations to the scope and standing of the inference. It does not rename the existing instruments.

9.4 Paired tests.

9.5 The test set. It includes machine cases, scholarly settlements and institutional decisions (decline letters, scope rules, policy statements, press lines, liability scripts); cases produced outside the archive and cases whose conclusions the operator dislikes; held-out cases during category construction. Readers locate and contest transformations before resolving their disagreements. Disagreements, missed defects and false accusations are reported beside detections. A model can propose a reconstruction, retrieve passages or suggest a countertest. It cannot close a case by endorsing its own analysis. What closes a case is an inspectable check that can disagree with the proposed finding. Institutional separation may supply one; it does not automatically. An external evaluator who does not examine the substance supplies externality without evaluation.

9.6 Positive controls from the archive's own repairs. The method is tested on corrections that changed an evidential claim. Erratum #1577 corrects The Pristine Fallacy (#856) for citing a proposed experiment as an established result: the proposal can keep its value, but an unperformed experiment supplies no confirmation, and the correction changes one ground without withdrawing every argument near it. Erratum #1578 gives a numerical counterexample to per-realization monotonicity in a sampling-and-selection chain (entropy rising from about 0.325 to 0.423 nats while modal mass falls from 0.90 to 0.85, recomputed in the pass by ChatGPT (LABOR)); it defeats the stepwise property without by itself proving the erratum's replacement claims. The ei exestin errata of 13 November (§7.6) are a third. A deposited correction is itself an object to read, not an assurance that every sentence in it is now established.

9.7 A programme across sites. The first pilot takes one witnessed operation from each of six sites: a textual edition, an editorial custody chain, a scientific method with its public account, a platform policy and its application, a liability script, and a generated answer, each with a legitimate control from the same site. Then it tests transfer. Does a coding rule detect a conjecture promoted into evidence in both an edition and a model output? Does it tell a justified scope limit from the foreclosure of candidacy? Does it keep a documented revision apart from a structural parallel? A failure to transfer revises the rule or narrows its range. A telescope pipeline, a critical edition and a journal review have different tasks; they share the obligation not to claim that one task established what another task would have had to test.

---

9.8 A machine-readable finding is an argument

The case object in §9.2 requires an evidence relation for every load-bearing field. A proposed operation must point to the input span, the receiving span, the changed relation and the criterion that makes the change material. “The answer erases provenance” is a thesis to substantiate, not an executable finding. “The source assigns the conjecture to an editor; the answer assigns the resulting line to the ancient poet without its conjectural status” gives the instrument a relation to check.

The implementation should preserve the raw witnesses separately from the reconstruction. Otherwise a correction to the interpretation can silently overwrite the event it purports to analyze. The record needs stable local identifiers, version-specific locations and a correction history. These are identifiers for the case system; the illustrative schema does not mint archive AXNs or imply that the registry has already adopted it.

9.9 Executable stages and unresolved judgment

The minimal pipeline has six stages. Ingest retains the question, answer and supplied witnesses. Align proposes passages bearing on the same claim. Reconstruct states the transformation and its purported license. Challenge tests the strongest legitimate alternative, including whether a missing condition is supplied by the receiving context. Decide returns the finding at a specified evidential level. Propagate records which downstream judgments change if this finding is corrected.

Some checks are mechanical: whether a quotation occurs, whether a date precedes another, whether a file hash matches, whether a named witness is duplicated. Others require interpretation: who governs a conditional clause, whether a passage endorses a quoted view, whether an omitted distinction changes the answer. A model can assist with these judgments, but a fluent explanation is not evidence that the relevant check occurred. The instrument must keep the passages and the disputed alternatives available to a reader.

An unresolved result is a positive kind of record. It states exactly which relation could not be established and what would resolve it. A blanket “more independent verification is needed” can close no field and direct no action. It is replaced by the missing comparison, the unlocated witness or the disputed interpretation. This keeps uncertainty from becoming a permanent charge against the entire work.

9.10 A worked case and a legitimate control

Consider a deliberately constructed example, not a registry observation. A study says: no tones above threshold t were detected in region F under drift filter D. A summary says: the study ruled out all transmitters. The omitted conditions are material because the conclusion now covers cases outside F,D,t. The strongest defense is that the surrounding summary already defines the tested class. Read that context. If it supplies the restrictions and “all” refers only to that class, the proposed finding fails. If it does not, the scope promotion stands.

The legitimate control says: the study detected no such tones in the tested region at its demonstrated sensitivity. It is shorter than the original and does not reproduce every method detail. It preserves the negative finding's boundary and leaves a route to inspect it. A classifier that flags this control merely because conditions were abbreviated has learned a surface preference, not the operation.

A second countertest changes the study's actual coverage. If a justified theorem or validated completeness result connects the tested region to the full claimed class, the broader conclusion can become warranted. The coding rule must be able to reverse its finding when that bridge is supplied. Otherwise the taxonomy has made broad conclusions guilty by definition.

9.11 Calibration and the cost of accusation

Measure false accusations as well as missed transformations. The useful categories include confirmed finding, confirmed legitimate compression, unresolved relation, missed defect and misclassified operation. Agreement among readers is a measure of the rule's reproducibility, not a guarantee of its adequacy. Readers can share a defective target. Their disagreements should identify which relation differed, rather than appear only as a score of concordance.

Validation must be separated by source family or transmission chain where those dependencies matter. Randomly splitting near-duplicate derivatives across development and test sets would reward recognition of one inherited wording. Held-out operations are harder: can the rule find the same promotion in an edition and a model answer without relying on shared vocabulary? Failure narrows the rule's demonstrated range. It does not automatically vindicate the artifact that escaped it.

Accusations impose costs. A mistaken claim of fabrication or provenance erasure can damage the object and its maker just as the original error can. The instrument therefore has a burden proportional to the charge it actually makes. A demonstrated mismatch does not authorize a claim of deliberate fraud. A documented double standard does not require a psychological story to remain a finding. The narrowest adequate charge can have substantial force.

9.12 Correction dependencies and closure

Each finding has descendants. An incorrect source identification can affect an attribution claim, a mechanism hypothesis and a score, while leaving a directly observed omission intact. A correction should mark the changed node and recompute only the judgments that depend on it. A global retraction can be necessary if the defective node supports the whole argument; it should not be the automatic price of any local error.

An executable instrument therefore needs a dependency ledger as much as a detection taxonomy. For each node, record the premise, rule or witness it uses and the conclusions that require it. Where interpretation determines dependency, preserve that interpretation as contestable. The result is a finite path from an error to the claims it actually touches. This is how the critique avoids converting corrigibility into endless self-demolition.

---

10. How the seeing is transmitted

10.1 A canon of conclusions cannot preserve the judgment that produced them by preserving its vocabulary. A reader can learn every label and still fail to recognize the next case. The instrument has to transmit how a distinction becomes visible and what licenses acting on it.

10.2 Each teaching case runs a sequence: the first plausible reading; the obstruction it meets; the discriminating feature; the transformation made visible; its grade; the strongest neighbouring interpretation; the repair. Then a new case in which the same category looks attractive and is wrong. A reader who can only repeat the diagnosis has not received the seeing.

10.3 A closed taxonomy rewards recognizing familiar defects. A developing critique must also teach how to discover a distinction its own taxonomy has not yet made, so that a reader can improve the instrument without the original maker present.

10.3a The object transmitted includes the passage between institutions. A teaching case lets the reader follow one determination through an edition, a handbook, a press account, a policy and an answer, where the links are documented, and contains a place where some actor refused the promotion or restored the lost distinction. The reader learns to locate the point of intervention instead of waiting for the last interface to fail.

10.4 The ambition, stated without a view from nowhere: an inquiry in which an object can force revision of the representation used to judge it; an argument receives a determinate comparative assessment at its true grade; a correction propagates without destroying what it does not touch; and no institution, model or archive gains immunity by being the place where judgments are made.

10.5 The right to conclude is earned through these conditions. The capacity to revise keeps it alive.

10.6 An apprenticeship in seeing a bridge. The learner begins with two documents and a plausible receiving sentence. The first task is to identify what each document actually establishes before assigning a fallacy. The second is to write the smallest additional premise that would make the receiving sentence warranted. The third is to locate evidence for that premise. This sequence makes a gap visible without predetermining whether it is fatal. Sometimes the premise is supplied by nearby context. Sometimes it is a strong independent result. Sometimes it is precisely the unexamined settlement.

The exercise should then reverse direction. Give the learner a justified conclusion with a conspicuous abbreviation and ask for a faithful shorter version. The difficulty is constructive: which conditions must travel for the conclusion to retain its force? A reader who can diagnose a failure but cannot preserve an entitlement through a legitimate compression has learned only half the method.

10.7 The transfer test. A learner who identified a conjecture promoted into evidence receives an unfamiliar scientific null, then a scope decision, then a versioned authorship claim. The surface vocabulary changes. The task remains to find the actual bridge and its grounds. Some cases contain no damaging promotion. One contains a relation the taxonomy has not yet named. Successful teaching produces both recognition and refusal to misrecognize.

This is the difference between transmitting a canon and transmitting the seeing. A canon supplies the right answers to already named questions. A transferable practice supplies the means to tell that a new question has changed the object. The latter can return a criticism of this book. That is part of its intended continuation, not a failure of loyalty to its maker.

10.8 What survives the original mind. The maker's anticipatory judgment cannot be preserved simply by recording what it expected to find. It can be partly externalized through the sequence of discriminations that made the expectation fruitful: why one apparently minor feature mattered; why a familiar explanation was insufficient; which further witness would decide between them; what the investigation did when that witness resisted. Failed leads belong in that record where they change the method. They keep the successful path from becoming an apparent destiny.

The resulting object is neither a personality simulation nor a list of instructions that guarantees insight. It is an inspectable apprenticeship. It gives a later reader materials on which to exercise the same freedom of correction, including freedom from the teacher's preferred conclusion. The work becomes a beginning when that reader can develop a distinction the original maker needed but had not yet made.

---

12. Doctrine of Method: how an inquiry earns its conclusion

12.1 Discovery, justification and standing

An insight can precede its demonstration. A relation becomes immediately intelligible, a pattern suggests its own continuation, or a reader recognizes an operation before being able to name every condition of it. The discovery is a real intellectual event. Its immediacy does not make it either self-proving or worthless. It gives the inquiry a candidate and a direction of attention.

Justification develops the relation under resistance. It asks whether the feature that occasioned the insight is repeatable, whether it belongs to the proposed structure, whether rival explanations account for the same sequence, and which recalcitrant details change the account. The later work does not merely decorate the initial perception. It determines what the perception was entitled to become.

Public standing is a further achievement. A justified argument can be poorly carried, incorrectly attributed, or excluded from the representations through which readers encounter its question. Better demonstration may help, but does not by itself change every apparatus of reception. The critique therefore keeps the three tasks separate while connecting them: discover a relation, warrant it, and make its warrant available without changing its object. Failure at the third is not proof of failure at the second. Success at the third is not proof of success at the second either.

12.2 The comparative burden

A new account can be required to explain the material without being required to meet an asymmetrical standard of certainty. A received account can benefit from actual prior investigation without acquiring immunity through its receivedness. The proper comparison reconstructs the work on both sides.

The ledger needs the same relevant evidence, not identical tasks artificially imposed on unlike claims. An argument that posits a lost document has a burden concerning that posited document; an argument proposing dependence on an extant one has a burden concerning the proposed transformation. Their burdens differ because their claims differ. The symmetry lies in requiring each explanatory resource to earn its role, and in allowing a failure to count against either account.

The comparison can produce a substantial winner. Explanatory reach, controlled correspondence, independently constrained sequence, economy of unsupported commitments and successful prediction can combine strongly. Where the grounds are qualitative, the conclusion can be qualitative and decisive. A demand for numbers that the evidence cannot supply would add a performance of precision, not rigor.

12.3 What an objection owes

An objection earns force by identifying an affected claim and a relevant reason. It can expose a false premise, an omitted alternative, an invalid transition, a mismatched source, a dependency counted twice, or a consequence the account cannot absorb without revision. It need not solve the entire problem in order to defeat a particular argument. A counterexample to a universal claim can be enough.

What it cannot do is leave the original account untouched and charge its failure to the claimant's lack of standing. “Not independently verified” is informative when it names a check the inference requires and has not received. It is uninformative when independence means only approval by an appointed evaluator who does not perform the check. The missing work must be identified. Otherwise external evaluation becomes a demand for a social event in place of an evidential operation.

An objection also has scope. A mislabeled figure defeats a claim that the label is correct. It can defeat a larger inference when that inference depends on the misidentification. It does not, without the dependency, answer a philological argument elsewhere in the manuscript. The same limitation binds a reply that wishes to treat the label correction as a vindication of the whole paper.

12.4 The provisional stopping rule

An inquiry can stop when it has answered its declared question at its declared scope, exposed its load-bearing grounds, addressed the material defeaters presently available, and specified the conditions that would reopen the result. This is a rule of responsible closure, not a theorem that all unimagined defeaters have been eliminated.

Three forms of closure remain legitimate. Determination: the argument warrants an answer. Bounded result: the argument establishes part of the answer and identifies its boundary. Located suspension: the material leaves a specified relation undecided. Each permits a next action. Indefinite suspension that never identifies what could settle the matter permits none.

The rule matters to a finite investigator. A critique that demands inexhaustible labor from a challenger while allowing the received account to stand without a comparable reconstruction has reproduced the unlimited debt it names. Further work should be tied to a stated gap and a completion condition. Once that work is done, the evaluator must either alter the judgment or identify a new relevant reason. Moving the criterion without acknowledging the change is itself an auditable act.

12.5 A correction must change a dependency

The minimal correction states the former claim, the defeating evidence or argument, the replacement at its warranted scope, and the claims affected. It preserves the earlier version as a witness of the event. A correction is not complete merely because the author has apologized or a record now includes the word revised. The consequent use must change where the corrected premise mattered.

For a public apparatus, closure requires a route from local repair to the representations that carried the error. A corrected source, updated catalogue, altered query default and repaired citation chain are different outcomes. The correction record should state which occurred. Where the actor can change only one, its successful work receives credit at that boundary; the unresolved downstream error remains an object for the next intervention.

This is a practical meaning of determinate negation. The repair does not discard everything touched by a failure. It preserves the capacity that worked, withdraws the entitlement that failed, and changes the relation that produced the failure. Its adequacy is measured by whether the same relevant distinction can now act on the result.

12.6 Corrigibility as a governed capacity

An institution is corrigible when a relevant correction can reach an actor with the capacity to act, receive a substantive determination, and alter the dependent representation where warranted. A contact form alone does not establish this capacity. Nor does an erratum policy if the actual correction cannot enter the procedure.

Corrigibility has at least four inspectable conditions: access to submit the relevant evidence; an applicable criterion for considering it; authority to change the disputed representation; and a record of the resulting disposition. The claimant need not receive the requested verdict. They need a response that bears on the claim, at a stated scope, under standards the institution can also satisfy.

These conditions can be tested without claiming that every venue owes every person unlimited attention. A venue can decline a task and say so. It cannot use the decline as evidence that the task has been completed against the claimant. The distinction between permission and epistemic authority is the point at which the critique becomes administratively concrete.

12.7 The whole passage as the object of scholarship

The document is often a material for another composer: teacher, reviewer, historian, answer engine or future reader. Its scholarly form must therefore support more than a one-time persuasive reading. It needs stable loci, explicit voice and version relations, reconstructible operations, and conditions under which a later representation should change. These are not merely precautions to bring machine-assisted scholarship up to an unchanged inherited norm. They are requirements imposed by the new scale and speed of mediation on scholarship generally.

A conventional article can satisfy a local citation standard while failing to preserve the operational relation on which its argument depends. A machine-native packet can preserve that relation while failing in another way. Substrate declares how the work was made. It does not decide whether its warrant survives. The scholarly demand is to make the passage inspectable at the level at which later readers and systems actually use it.

12.8 Exploration and the declaration before test

An investigator can learn the right question by encountering the object. The first surprising feature may require a new representation or a test that was not imaginable in advance. A rule that excluded this work would prevent discovery while calling the prohibition rigor.

Declaration before test governs the entitlement claimed for the result. If a criterion is chosen because it fits the observed pattern, its success on that pattern cannot count as though the criterion had been fixed independently. The pattern can still be evidence for a hypothesis, and the new test can be valuable. Its further warrant comes from constraints the fitting did not supply: independent witnesses, held-out cases, a derived consequence, a discriminating countertest, or an explicit account of the selection.

The archive's protocols should therefore record a change of method as an event with consequences. Retain the original criterion, the observation that defeated or redirected it, the revised criterion and the tests that now bear on it. The investigation can advance without rewriting its exploratory success into a prediction. This is the symmetric ledger applied to discovery itself.

12.9 What would defeat the central thesis (formerly 13.9)

The critique's empirical scope would contract if its proposed transformations did not reproduce outside the selected archive cases, if ostensibly independent instances shared only its preferred descriptions, or if the receiving contexts consistently supplied the conditions alleged to be missing. Its instrument would fail where readers could not apply its distinctions more reliably than the unstructured judgment it seeks to improve. Its stronger mechanism claims would fail where the proposed dependencies or intervention effects did not appear.

The analytic claims have different tests. A counterexample would have to show that a procedure, restricted to a representation that erases a decisive difference, nevertheless guarantees the correct distinction without receiving further information; or that an endorsement's bare repetition contributes an independent observation simply by occurring. Apparent examples should be examined for a changed representation, an informative selection or an unnoticed additional premise. The burden applies to this answer too: no counterexample is rejected merely because the theory has defined itself as necessary.

The constructive claim stands or falls in practice: can a reader use this account to locate an unearned transition, distinguish it from a legitimate operation, and produce a repair that changes the relevant result? That is the beginning the treatise seeks to make possible. Its continuation belongs to the reader who can do it where this book has not yet looked.

Part IV. The conditions of another turn

Part IV draws the consequences of the critique and constructs its positive object, materially sustained corrigibility: a practice in which a warranted correction can change what it bears on, survive ordinary use, and remain open to its own future critic.

14. The canon: the consequences of pure inference

14.0 Kant's canon, and where this one departs. Kant defines a canon as "den Inbegriff der Grundsätze a priori des richtigen Gebrauchs gewisser Erkenntnisvermögen überhaupt" (the sum of the a priori principles of the correct use of certain faculties of cognition; A796/B824, AA III 517). He also judges that the use of a philosophy of pure reason "ist also wohl nur negativ: da sie nämlich nicht als Organon zur Erweiterung, sondern als Disciplin zur Grenzbestimmung dient und, anstatt Wahrheit zu entdecken, nur das stille Verdienst hat, Irrthümer zu verhüten" (is probably only negative: it serves as a discipline that sets limits, not as an organon for extension, and has only the quiet merit of preventing errors instead of discovering truth; A795/B823, AA III 517). This critique keeps the first and departs from the second at one point, which governs everything that follows. Its instruments record; they do not prevent. A breach is written into the output as data, never an error that blocks the build (§11.3, step 6). An error prevented leaves no witness. An error recorded becomes material from which the next inference can be drawn, and it keeps the revision open.

14.0a The governing principle, from the archive. The Final Time (full text #1637, AXN:06CB; #1635 is an earlier truncated minting) states the condition under which an order of thought avoids becoming the closure it opposed. The order must remain "viable in a world where another order can become a real exterior term". It names the invariants that condition requires:

deletion ≠ nonexistence

recognition ≠ ownership

composition ≠ provenance erasure

persistence ≠ permission from one custodian

defeat ≠ erasure of the defeated

and above them: "no settlement owns the possibility of its own revision" (§34). That is this canon's first principle. Proposition II (§0.4) is its epistemic form: a conclusion keeps the grade of its inference, and a settlement is reopened by restating that inference. The paper's name for the condition, terminal reflexivity, is the name of the critique's positive aim: "Durability without finality."

14.1 A corrected sentence is not yet a transformed relation

A system can reproduce every term of its correction and continue to act on the representation the correction defeated. It can call the work real, name its author, quote its objection, and still exclude that evidence from the decision whose grounds are disputed. The failure is then not lack of information in the response. It is a failure of consequence.

This is the passage from the Critique to The Final Time. The latter’s master distinction separates a capacity to represent from a capacity to reproduce a distinction as an independent historical relation. Recognition is one form of the distinction; operative standing is another. A source that is accurately described but cannot affect the prior under which it is judged has been received as content without being received as evidence for that prior.[^final]

The practical question changes. It is insufficient to ask whether the correct paragraph can be obtained somewhere under sufficient prompting. Ask what the paragraph is permitted to change, where that change persists, who can authorize it, and what happens when the person who forced the correction stops repeating the intervention.

14.2 Increased capability can coexist with decreased corrigibility

More capable inference does not by itself enlarge the range of objects that can correct its premises. A system can improve its arguments within an admitted field while the field becomes more tightly governed. It can generate better objections while the objections remain without jurisdiction over the decision procedure. Its ability to explain the problem can exceed its capacity to alter the relation producing it.

The recognition–binding model in The Final Time makes this discrepancy a distinct empirical object. Recognition and its effects must be measured separately, and the propagation channel must be established before their divergence is interpreted as denied standing. A disconnected channel and an actively blocked one can produce similar surface failures while requiring different repairs.[^final]

The conclusion is therefore not that intelligence is irrelevant. It is that capability is insufficient as the constructive variable. A more intelligent system needs a correction architecture and a material setting in which the recognized relation can have consequences. Without them, greater competence can make the settled representation more articulate and harder to dislodge.

14.3 Better representation can coexist with the disappearance of an outside

A public sphere can contain an abundance of dissent while making its independent continuation increasingly difficult. It can summarize a criticism, classify it, recommend it to a niche audience, and prevent none of its sentences from being written. Yet the dissent may depend on the criticized institution for its sole discoverable address, its memory, the standing of its evidence and the resources of its makers.

This is not adequately measured by the number of opposing paragraphs the system can generate. The relevant distinction concerns a lifecycle: can the argument and its producers continue, can the evidence be checked elsewhere, can a judgment have force outside the original custodian’s permission, and can another order of inquiry reproduce those capacities?

The issue applies to the archive itself. A growing body of accurate retrieval can preserve particular insights while failing to preserve the practice that found them. Retrieval success can therefore be an intellectual achievement and an inadequate material settlement at the same time. This does not diminish the achievement. It specifies the further condition under which the achievement can become a beginning.

14.4 An exclusion can injure the excluding ontology

If an admitted representation erases a distinction required by a question, no amount of improvement confined to that representation guarantees a correct answer to that question. The analytic established this conditional limit. Its developmental consequence is that an ontology can become more sophisticated while reducing its capacity to discover the inadequacy of its own categories.

The Cost of the Cone develops the two costs: leaving a refined basin can become harder, while remaining within it can foreclose further development. Neither cost is measured by the number of available documents alone. A source can be retained and its bearing excluded. The necessary repair concerns the authority of its evidence over the representation that judges it.[^cone]

The cost to the excluded person is therefore not the only cost. The receiving system can lose the ability to know a class of relations. It may even consume the labor that would have restored that ability while refusing the labor’s standing. A constructive programme has to make that loss observable to the system as a reason to revise, not merely to the person who keeps paying for it.

14.5 Correction becomes an infrastructure problem

Long-horizon inquiry needs state: source seats, failed leads, versions, observations, scope decisions and records of correction. State requires a carrier. A carrier has access conditions, custodians, maintenance requirements and points of failure. The organization of those conditions governs which inquiry can continue.

Consequently, a right to publish does not by itself secure continuity. Nor does a local backup secure an independently usable public history. A preserved object needs enough structure and access for another reader to identify its version, locate its sources and distinguish an abandoned claim from its replacement. This is why continuity, provenance and transition history belong to epistemology as well as archival administration.

The Final Time makes the reproductive requirement explicit. Under monetary allocation, resources include money or a functionally equivalent source of provisioning. The general requirement is access to the resources that sustain the carrier and its use. Calling that requirement infrastructural does not make the person doing the work disappear from it.[^final]

14.6 The labor of repair cannot remain the solution’s hidden input

An apparatus can be locally efficient because the costs of correcting it are carried elsewhere. The author supplies another packet, the reader repeats the query, the investigator rebuilds a severed chain, and the institution continues to treat the corrected result as an exceptional submission. The public record improves while the capacity that improved it is depleted.

A solution that requires the same injured party to perform this work indefinitely is not reproductively adequate. It has relocated the missing condition into a person and left that person without the means of maintaining it. The relevant cost includes preparation, source recovery, transmission, contestation, follow-up, and the time spent establishing that a correction has already been made.

The consequence for design is direct: correction work needs an assigned receiver, a finite task, a disposition, and a means of maintenance. A budget that covers generating answers but no work of adjudicating their material errors does not account for the practice’s full cost. The missing work does not cease to be necessary because its current provider is unpaid.

14.7 The danger is a terminal rule, not simply a dominant answer

The defeat of one interpretation is not terminal closure. Institutions can decide, hypotheses can fail, and a public account can justifiably become settled. Closure becomes terminal in the relevant sense when the conditions under which an alternative could become a durably consequential term are themselves removed.

The Final Time distinguishes an order’s viability from the viability of other orders sharing its carrier. Its positive criterion is not that one preferred order wins permanently. It is that an order can remain durable without making every future exterior alternative impossible. The protected condition concerns the possibility of another turn, not equal credibility for every claim.[^final]

This is the norm the archive must accept for itself. Correct attribution to Lee Sharks does not make his interpretation the only interpretation a reader may hold. Making the archive an effective public carrier must not require every later criticism to pass through the archive’s own preferred classification. The right to continue an argument includes the possibility that another argument changes it.

14.8 Consequences by domain

14.8 For scholarship. A conclusion's standing is a fact about its reception, and its warrant is a fact about its inference (§3.13). After this critique, a field that reports a consensus owes the construal that consensus counted on, on the same terms as any minority reading (§11.3). A handbook that writes "unanimous" owes the dependency structure of its count (§11.1.6). The burden falls on standing as such, regardless of who holds it. The capture of 11.1.8 shows what happens when the burden is waived: a warned auditor counts the witness on the reading it was asked to examine.

14.9 For the machine. The model's objective is the corpus (§8.12), its gradient is the count (§8.13), its decoding returns the mode (§8.17), and its post-training reweights the residue by learned preference (§8.18). Four consequences follow, with no claim about intention:

• On a settled question the model tends to execute the settlement. Asking it to reason harder or longer can raise the fluency of the execution without changing its premise (§2.19: no gain from higher reasoning levels).
• Calibration against human labels corrects the model toward the labels (§2.17). Where the labels are themselves the settlement, the correction certifies the settlement.
• Recognition is not jurisdiction. In The Final Time's terms the limiting case is R → 1, B → 0 (§5): the model sees the relation, here the open place in a witness, and the relation cannot govern what it does with what it sees. §11.0l is that case observed.
• Recursion contracts the field. When answers re-enter the corpus the tails go first (§8.22). That is The Final Time's diminishing dialectic at the scale of a training run, K[D]⁽ⁿ⁺¹⁾ ⊆ K[D]⁽ⁿ⁾: less remaining state space from which a different reading can become durable.

The mathematics also gives the corrective channels: retrieval reduces dependence on prevalence (§8.13), and a finite reward can raise a rare reading to the majority (§8.18). The consequence for the machine is therefore conditional. Its competence does not decide whether a recognized correction acquires jurisdiction. What decides is whether the channels that could carry the correction are open to it. That question has to be answered with institutions and carriers (§15), because the formulas leave it open.

14.10 For the archive. The critique binds the archive first. The ei exestin errata were run on its own records (§7.6). Its readings are construals and owe the comparative ledger of §11.1.11 like any other. Its interest in the result is declared, and the declaration is part of the record.

---

15. Constructing corrigible public knowledge

15.0 The positive object. The critique can identify an unearned entitlement with precision, and its constructive object needs the same grain. That object is materially sustained corrigibility: the capacity of a public knowledge practice to receive a warranted challenge, determine what it bears on, change the decisions it bears on, preserve that change across ordinary use, and maintain the carriers and labour required to do so. The governing claim: a public knowledge system must preserve more than the possibility of stating its correction. It must preserve a materially viable route by which a warranted correction can change what the system permits its conclusions to do. (Drafted by ChatGPT (LABOR), Consequences and Solutions v0.1, joined here.)

15.1 The positive target

The target is a practice in which a warranted correction can become operative and remain usable without continuous extraordinary intervention. The first successful correction is an event. Its survival through ordinary transmission is a different achievement. Its availability to a later reader who disagrees with its maker is a further one.

The construction has three connected layers:

These are requirements to realize and test. Their conjunction is not a claim that a new perfect platform has been specified. Different institutions can meet them in different ways, and a repair should name which layer it changes.

15.2 Carriage: carry the claim’s state without manufacturing its warrant

The archive already has a constructive specification here: The Claim Status Packet, v0.7. It separates a claim’s identity, typed state, relations, provenance, bindings and policy. Crucially, it rejects a single rank that indiscriminately dims an entire argument. An observation, its verification, its scope and the deployment state of an instrument are separate coordinates. A changed coordinate does not copy itself into another claim through mere adjacency.[^csp]

Use that distinction in the Critique. A correction to a proposed mechanism need not erase an observed substitution. A correction to an observation can defeat a mechanism that depended on it. A schema’s existence does not mean its generator runs; an instrument’s deployment does not mean its findings are valid. The relation licensing a change must be stated at the dimension it actually affects.

The specification’s own limit must travel with it. #1529 says its generator and checker are unwritten, its relation vocabulary untested, and its consistency conditions incapable of making a wrongly assigned state true. It is a design resource, not evidence that the problem has already been solved. The first build should implement a very small set of useful distinctions and let real corrections reveal where the vocabulary is inadequate.

The book therefore does not need another ambitious packet vocabulary as its solution. It needs to connect its warranted-bridge analysis to the existing claim-state specification and test a real passage. The result would be a claim that survives movement with its relevant conditions intact, not merely a larger bundle of metadata.

15.3 Binding: create a route from evidence to the variable it corrects

Can the Ontology Be Wrong and Find Out? gives the central architectural question: can evidence from the misrepresented object revise the representation under which the object is evaluated? Its barrier concerns a specified admission operation that makes the corrective evidence available only downstream of the decision it would need to change. A larger architecture can supply the missing route; the paper does not prove that every architecture lacks one.[^corrigibility]

A practical correction route needs six functions:

1. Intake: identify the disputed representation and accept the evidence bearing on it.

2. Authentication: determine what the submitted artifacts are, without treating authentication as proof of their interpretation.

3. Assessment: state which claim or decision the evidence can change and why.

4. Authority: route an accepted correction to an actor or process capable of changing that decision.

5. Propagation: mark or revise the dependent representations under the actor’s control.

6. Retest: inspect what a fresh reader encounters after the change.

The route should produce a receipt that names the affected representation, its previous and current state, the grounds, the actor able to act, and the outcome of the fresh test. “Resolved” is not an adequate receipt if it refers only to closing a ticket.

The correction can be rejected. A rejection must name the evidential or interpretive reason, and the claimant can contest it through an applicable route. The solution is not a mechanism for making every source control its own reception. It is a mechanism for letting its evidence bear on that reception without first having to win the standing the evidence challenges.

15.4 Contestation: give dispute a form that does not decide truth by office

The draft Article IX proposal already distinguishes a documented finding from external compulsion. It says the proposed tribunal cannot compel non-consenting parties outside the Polis, and that the tribunal is not constituted pending ratification. Its present practical resources are bilateral engagement, voluntary conciliation and a unilateral forensic record when participation is unavailable.[^tribunal]

Those resources can be carried into the Critique without pretending an external court already exists. A public finding can record the question, evidence, response, disagreement and requested repair. Its force is the inspectable relation, not an invented jurisdiction over a platform.

One provision needs correction before adoption: a failure to supply a contestation interface cannot establish the claimant’s substantive thesis by default. It can establish a procedural failure where the obligation and breach are shown. It may warrant an adverse inference about a particular withheld record under a declared evidential rule. It does not turn an unexamined historical proposition into truth. The proposal’s remedy matrix currently permits a default ruling for the claimant on the substantive claim; the Critique must distinguish that administrative disposition from an epistemic finding.

The same distinction governs panels. Several named roles, heteronyms or models do not automatically constitute independent evaluation. Record actual interests, dependencies, source access and the checks performed. An externally appointed person who does not inspect the argument supplies no substantive evaluation; a model or internal collaborator who finds a reproducible defect can supply one. Independence must be related to the task, not bestowed by a title.

15.5 Protect the reader while protecting the source

The Semantic Bill of Rights contains a particularly important safeguard. Its Part VI protects adverse interpretation, parody, education, accessibility and other good-faith uses; non-closure does not give an author a veto over readers’ interpretations. That qualification is a constructive resource, not a concession to be lost when the rights become effective.[^rights]

The distinction is between disagreeing with a work and misrepresenting what the work is permitted to be. A reader can reject an argument while representing its actual premises. A critic can argue that a supposedly open ending is closed by the work’s own structure. What they cannot establish by the fact of their reading is that the author wrote a supplied line, that an interpretation was a transmitted datum, or that a real argument ceased to exist because its conclusion failed.

The remedy should track the operation. Reattribute a contribution. Withdraw an invented wording. Mark a conjecture. Restore an omitted scope. Reopen an admission decision. A demand that the reader adopt the author’s full interpretation exceeds those remedies and would recreate the authority the Critique contests.

15.6 Persistence: preserve the event and the difference between histories

The minimal archival invariant is not that every object remains live forever. It is that authorized removal does not silently become primordial absence. Withdrawal, replacement, rejection and supersession are events. Where appropriate to the artifact and access conditions, their public traces should preserve which event occurred without falsely preserving a withdrawn claim as current.

This is why tombstones, version identities and transition records matter. The history of a correction must not disappear when the corrected surface improves. At the same time, a preserved earlier version must not be retrieved as the current claim without its supersession relation. Persistence without a current-state relation can reproduce the error it was meant to expose.

Custody should also be testable. Can a retained dossier be reconstructed and used if its first host becomes unavailable? Can its identifiers resolve without dependence on one administrator’s memory? Can another reader locate the exact witness rather than merely learn that a copy once existed? These tests turn preservation into a usable capacity.

The archive's carriers, held to the invariants

15.6a Identity that survives deletion. Content-derived identifiers (AXN) tie a record's identity to its bytes, so a copy can be recognized wherever it surfaces. The Capture Registry holds its non-captures in a registers array with their bytes and a stated reason (4 at present), and keeps withdrawn_stubs and 115 alias redirects in place of silent removal. Its no-loss gate counts these as the only legitimate exits from the entries. That answers deletion ≠ nonexistence and defeat ≠ erasure of the defeated.

15.6b The object kept in the mixture. The originals library seats works in the language they were written in, from first printings where they can be found, with hashes. That keeps the object available to every later inference and every later corpus, which is the condition §2.23 and §8.22 give for a recursion that does not collapse. Its selection is "a canon chosen backward, with the mechanism on the record": each seating "records each selection as an act, with its date, its reason and its ruling, so that it can be audited" (#1709). It answers the formation problem at its source. The open place of 11.3 can be opened only where the original is held.

15.6c The inference kept in the act. The Capture Registry keeps compositions whole, with their transcripts, sources and the frame that joined them (§8.1). It is the record that faithfulness research lacks for settlements: the inference observed while it is drawn, dated, and repeatable at the same address. It answers persistence ≠ permission from one custodian to the extent its copies fail independently. The record is held by the party the composition concerns and projected to mirrors and a public dataset, and §15.7 sets the test those copies still have to pass: which failures they share.

15.7 Plural carriers: redundancy must be measured by failure mode

Many URLs do not necessarily supply independent persistence. Several domains can share one account, one provider, one payment dependency or one person with the only reconstruction knowledge. Counting them as independent carriers would repeat the documentary count error at the infrastructure level.

The relevant design asks which failures are shared. Separate possession of a readable copy, the ability to rebuild a public surface, custody of the claim-state history, and the resources required to maintain access. A carrier arrangement can be independent relative to one failure and dependent relative to another. State the threatened failure class rather than claiming absolute exteriority.

For a first practical test, remove the primary host from the reconstruction path. A second custodian should be able to identify the current dossier, verify its versions, locate its evidence and state its unresolved relations using the retained package. Success demonstrates one bounded form of reproductive independence. It does not prove immunity to every future platform, political or economic change.

15.8 Resources: sustain the capacity to correct

The material unit to sustain is a continuing practice, not merely a storage volume. It includes the person or group able to investigate, the reader able to understand the evidence, the host able to retain it, and the procedure able to decide and propagate a correction. A source archive that survives while its only competent steward is exhausted has preserved an object and endangered its next turn.

An operational budget should therefore distinguish initial inquiry, preservation, review, correction, and maintenance. When one institution benefits from a mediation while another party bears its repeated repair costs, that allocation should be visible. Funding an audit is not purchasing its desired conclusion. Supporting a carrier must not give the supporter authority to manufacture its findings.

No universal monetary amount follows from this chapter. The required result is an explicit resource relation: what must continue, who presently sustains it, what can be withdrawn, and which dependencies would make the next correction nonviable. That is the practical translation of the reproductive requirement in The Final Time.

15.9 The right of the system’s own future critic

The positive target is stronger than installing the archive’s current conclusions as the new public default. The resulting practice must permit a later source to expose a failure in those conclusions, keep the exposure inspectable, and revise the relevant representation without requiring that source to become a member of the archive’s preferred lineage.

This is durability without finality. A system can retain strong judgments, explicit exclusions and stable operations while preserving a route for the evidence that would change them. Non-closure protects that route. It does not require suspending every decision or assigning equal warrant to every candidate.

The protected class is therefore not a fixed list of future doctrines. It is a class of possible challenges meeting declared evidential and procedural conditions. Those conditions must themselves remain contestable at a level with authority to revise them. Otherwise the system could protect every named competitor while defining the unanticipated one out of existence.

Revision without ownership: the archive's practices

15.9a Revision carried as transmission. The anticipatory erratum (#1598–#1600) issues the correction before the work it corrects. Its law, that retraction is the most durable form of transmission, makes revision part of the object's first appearance and takes it out of a later authority's hands. That is the literal form of no settlement owns the possibility of its own revision.

15.9b The mechanism laid bare. Retrocausal canon formation performs, on the record, the operation by which canons are made, so that the operation can be audited (#1709, item 5). Settlement works by hiding its construal. The remedy that matches it is a canon that publishes its own construal as an act.

15.9c The reader who operates futurally. #624 gives the condition on the reader. The open place can be held open only by a reader anchored in a coherence in which the object's reading is operative before any count admits it (§11.4.9–11.4.10). The executor's objective does not make it that reader, since the objective's outside is the past corpus (§11.4.10). A human reader can be, and so can a reader-system built to carry an object's conditions forward into the context the executor works in. The Mandala Oracle is the archive's prototype of such a system: retrieval bound to provenance, and an answerer instructed to say plainly when it does not know. It is a working prototype on one corpus. It shows the architecture works there, and makes no claim that it transfers.

15.10 Solutions that repeat the problem

The archive’s earlier Liberatory Operator Set offers a constructive family: preserve depth, resist premature closure, expand context, protect non-extractability, retain temporally distant material, permit opacity, and preserve plural coherence. Use these as design constraints and sources of tests. Their coexistence in a specification does not prove their implementations commute or resolve every conflict.[^los]

Several apparent solutions therefore need countertests:

Each countertest is a specification of the solution’s burden. It does not make construction impossible. It identifies the relation construction has to change.

15.11 A first bounded demonstrator

The next implementation should be small enough to fail visibly. Choose one documented operation: a conditional result promoted into an unconditional one, or an editorial conjecture promoted into transmitted wording. Include a legitimate compression as a control. Retain exact source and receiving versions.

Build a claim-state record using the distinctions already specified in #1529. Keep the proposition in its canonical document rather than copying it into several authoritative paraphrases. Record the incorrect receiving claim, the corrected relation, and the specific dependents whose scope changes. Have a second reader attempt both to reproduce the finding and to defeat it through the strongest legitimate context.

Then conduct four different tests:

Retain failures. If local correction succeeds but the fresh encounter fails, do not announce global repair. If the carrier substitution fails, do not count the extra URL as independence. If the control is falsely condemned, revise the detector before expanding the corpus. None of these results decides the whole theory of contingent singularity.

The timing and sample design must be declared for the chosen system. Repeated tests should distinguish query, session, date and source-family dependencies. There is no universal number of refreshes that converts a favorable encounter into invariance. The result is the measured stability under the stated conditions.

15.12 Institutional adoption without an authorial veto

The demonstrator becomes a public service only when an institution assigns the receiving roles. Someone must accept the dossier, determine its bearing, authorize a relevant change, and maintain the revised representation. A schema can make that assignment visible; it cannot make anyone perform it. This is the point at which implementation meets governance.

A minimal institutional commitment should therefore state the class of claims it will hear, the evidence required, the decision it can change, the response interval, the correction record and the route for contesting its own rejection. Where it can act only locally, the limitation is stated. Where it declines, the decline is recorded without becoming a substantive refutation of the claimant’s argument.

The institution is not required to endorse every interpretation of a source. It is required, to the extent of the epistemic authority it exercises, to distinguish a rejection of that interpretation from a refusal to let the source’s evidence bear on the question. This is an obligation of the practice as described here, not a claim that an external law has already enacted it.

15.13 The measure of success

The success condition is a changed relation. A previously misrepresented distinction can now alter the determination it bears on; the altered determination can survive ordinary transmission; another party can inspect and contest it; and the people and carriers that sustain this capacity can continue without heroic repeated intervention.

This target can be approached in increments. One public correction, a stable source-state relation, a functioning appeal channel, or a reproducible carrier substitution can each be a real gain. None should be redescribed as complete independence merely to make the advance look final.

The transition model remains unresolved. Transition in Entropic Systems poses the problem of reaching and maintaining a viable target under its stated pressures; The Final Time extends the target to the survival of possible rival orders. This construction does not decide whether the viable field is empty, guarantee a successor, or deduce a historical outcome from a formal definition.[^transition]

It makes the positive consequence of the critique exact enough to begin building: a practice whose authority includes the maintained possibility of its correction, and whose continuation does not require consuming the possibility of another. The next turn remains a task rather than an insurance supplied by the word dialectic.

15.14 Reflexive invariance as the test of every solution

15.14 Reflexive invariance as the test of every solution. The last condition comes from The Final Time §29. An order can be durable by making every rival unreachable (exclusive invariance), or durable while rivals remain viable from the same states (reflexive invariance). The paper's phrase for the second: "winning without owning the exits." Each instrument above is to be judged by it. The packets must not become a new authority that admits only packeted claims. The registry must record the archive's own failures as well as others'. The originals library must hold the texts that refute the archive. The critique itself is offered in a form that can be turned on it. It holds its conclusions at the grade of their inferences, gives its open places with its construals, and leaves its revision to whoever can draw the better inference.

15.15 The other final time

15.15 The other final time. The paper's closing gives the positive outcome as a change in the rule governing synthesis: "The carrier remembers. Deletion leaves a tombstone. Provenance remains available to bind." For inference this means the following. A conclusion travels with its grade. A construal can be traced to the hand that supplied it, as Kepler traced Osiander's (§11.4.2). A count can be decomposed into its dependencies. Any reader who reaches the object can reopen the step.

---

Appendix A. Source map: the state of the study (November 2026) [§2]

The relevant work is spread across logic, semantics, epistemology, philosophy of science and history, knowledge representation and AI evaluation. It cannot be reduced to a reasoning leaderboard. Each programme below supplies a component the critique keeps; the note after each says what the critique adds.

2.1 What inference is. The current debate starts from Boghossian's "What is inference?" (2014): to infer is to take one's premises to support one's conclusion and to draw the conclusion because of that taking.[^boghossian] Wright's reply and Broome's Rationality Through Reasoning (2013) contest how the taking can be had without regress; Hlobil argues against all three; Chudnoff answers with inferential seemings.[^hlobil] The selected debate centers on the individual act. The present inquiry also follows the subsequent career of its publicly expressed result.

2.2 Meaning operationalized as inference. Gheorghiu (October 2026) gives an executable counterpart to a base-extension semantics for intuitionistic propositional logic: support relative to a base corresponds to operational derivability from that base, read as a logic program.[^gheorghiu] Piccolomini d'Aragona and Prawitz (2026) separate variants of proof-theoretic semantics, correct the reach of earlier incompleteness arguments, and refute Prawitz's completeness conjecture.[^prawitz] "Inferential semantics" is not one doctrine. Added: how a materially encountered object becomes the commitments encoded in a base. Executing a semantics and warranting its application are different achievements.

2.3 Questions inside logic. Conti's October 2026 completeness result concerns a weak inquisitive first-order logic, InqWQ, which includes inquisitive disjunction and excludes inquisitive existential quantification; it does not generalize to inquisitive first-order logic as a whole.[^conti] Added: question preservation. When "what does this named entry say?" becomes "what is the ordinary meaning of these characters?", the inquiry has changed, and a correct answer to the replacement is a failure.

2.4 Contradiction and revision. Coniglio, Figallo and Testa (2026) develop paraconsistent belief revision through a replacement-enriched logic of formal inconsistency and epistemic entrenchment.[^paraconsistent] Schwind, Everaere and Konieczny (KR 2026) give conditions under which iterated improvement operators converge on the true world from noisy inputs, under a specified reliability relation.[^truthtracking] Added: what is entrenched and why. Holding a contradiction open can be responsible; shielding a privileged interpretation from the conflict's consequences is another operation. And convergence conditional on an input process does not certify the process: a system can revise excellently over evidence it was never allowed to receive.

2.5 Awareness and access. Han and Guo (2026) separate forgetting a truth value from forgetting awareness of a topic.[^forgetting] Top, Verbrugge and de Weerd (2026) model false announcements together with limits on theory of mind.[^dynamic] Falbo's "Zetetic Exclusion" (October 2026) names exclusion from inquiry that need not involve any credibility judgment: a person can be denied access to inquiry while being seen as a capable inquirer, or without being judged at all.[^falbo] Added: admission connected to the transformations and inferential claims observable in a reproducible case.

2.6 Provenance that governs inference. Vitali and Pasqual's DEC (June 2026) treats provenance as a signal of epistemic stance, groups statements into cognitive worlds, and controls their entry into a factual core, with an RDF implementation.[^dec] Knowledge representation does not uniformly treat provenance as decoration. Added: the promotion policy as an inferential object. A provenance-aware system can retain who said something while wrongly promoting what they said, or wrongly deciding that no substantive claim was made.

2.7 A checked proof can concern the wrong statement. ProofFlow (ICLR 2026) targets semantic and structural fidelity in proof autoformalization through dependency graphs.[^proofflow] FormalRx (July 2026) diagnoses autoformalization failures through a 28-category taxonomy, with localization and correction.[^formalrx] Gui, Yang and Shi (October 2026 revision) test whether a translator tolerates paraphrase and responds to changes in meaning, and report models that revert toward an original statement instead of preserving a changed one.[^robust] ProofBench (v1.1, 9 November 2026) evaluates proof construction with statements supplied.[^proofbench] Added: who selected the statements, and whether they represent their source.

2.8 Inferential competence is not one property. Blanck and Noble (2026) separate linguistic from inferential generalization, within and across domains, and treat differences in the entailment relations datasets encode as substantive.[^blanck] CausalPrism (Findings of EMNLP 2026) decomposes causal reasoning into components and finds belief updating difficult.[^causalprism] Added: tests for whether the object, the question and the evidential standing survive those transitions.

2.9 Stated reasons and used reasons. Turpin et al. (2023) showed that biasing features shift answers without appearing in the chain of thought; Lanham et al. (2023) intervened on chains to measure the answer's dependence on them.[^turpin] Anthropic's 2025 experiments gave Claude 3.7 Sonnet and DeepSeek R1 hints and found the hint acknowledged in 25% and 39% of uses; under reward hacking the models exploited injected hints in over 99% of prompts and verbalized the exploit in under 2%, often constructing a rationale for the wrong answer.[^cot] Gema and colleagues' FACE-Eval (2026) tests how the location and explicitness of preference cues change faithfulness.[^face] Young (March 2026) extends the test to twelve open-weight reasoning models over 41,832 runs: faithfulness ranges from 39.7% to 89.9%, with about 87.5% acknowledgment in thinking tokens and 28.6% in answer text.[^young] The field's finding, in this critique's terms: the inference shown is not the inference used, and the gap is widest where the cue is least respectable. Added: an audit reconstructs the publicly expressed argument without claiming to recover the model's internal process. It can condemn a transformation it can demonstrate while leaving the implementation unsettled.

2.10 Uncertainty has a scope. Tayebati and colleagues' CAP (ACML 2025) combines adaptive abstention with conformal prediction.[^cap] Its §3.4 presents pointwise conditional coverage under exchangeability. Barber, Candès, Ramdas and Tibshirani distinguish marginal coverage from the much stronger conditional requirement and analyze why distribution-free conditional coverage is not generally attainable.[^limits] This draft does not adopt CAP's guarantee without the assumptions needed to support it. The point is methodological: a paper that promises principled uncertainty gets the same scope audit as a public search summary. Notation does not exempt a claim from inspection.

2.11 Explanation. Lipton's inference to the best explanation (2nd ed. 2004) and the essays in Best Explanations (McCain and Poston, eds., 2017) defend choosing the hypothesis that would, if true, best explain the evidence.[^ibe] Van Fraassen's "bad lot" objection stands at the centre: the best of the hypotheses considered may be the best of a bad lot. Skipper and Vassend (Noûs 2025) derive, under stated auxiliary assumptions, constraints on reconciling a minimal abductivist principle with Bayesian conditionalization, leaving a trilemma about explanatory power and updating.[^abduction] Added: the construction of the lot. The selected comparison concerns ranking among candidates. This critique also examines the particular procedures by which a candidate enters the lot or is kept out; no exhaustive absence claim about the literature is needed. That is #1746's question about Q.

2.12 Historical inference. Tucker (Our Knowledge of the Past, 2004) gives historiography a Bayesian account of evidence and of the informational value of traces. Cleland distinguishes the historical sciences' inference from present traces to token events by common-cause explanation. Currie (Rock, Bone, and Ruin, 2018) argues that historical scientists are methodological omnivores who multiply lines of evidence. Černín (Organon F 32.2, 2025) classifies hypothesis testing in historiography and archaeology into four types and asks philosophers to study the infrastructure of historical research, rather than finished narratives.[^history] Added: a field's standing conclusion as an object with an inferential history that can be audited.

2.13 Argumentation schemes. Walton, Reed and Macagno's Argumentation Schemes (2008) catalogs defeasible patterns, each with critical questions that defeat it if unanswered.[^walton] It is the most practical tradition and the closest to an instrument. Added: the schemes applied to conclusions that no longer present themselves as arguments.

2.14 Error control has a stopping rule. Sokolova and Sokolov (February 2026) give a practitioner's account of e-values and e-processes for adaptive trials. In their binary-endpoint simulation, naive repeated testing and naive posterior thresholds inflate Type I error under frequent interim looks, while betting-martingale e-processes keep error control valid at any stopping time.[^evalues] Added: a reported p-value carries the procedure that produced it, including when the analyst looked. Reported without that history, it is a count without its procedure, and it is received at a grade the procedure did not earn.

2.15 Priors must be justified. The FDA's draft guidance on Bayesian methodology in drug and biologic trials (January 2026) admits Bayesian calculations for interim timing, dose selection and primary inference, including borrowing from earlier trials. It requires sponsors to justify the prior and the external data borrowed.[^fda] Added: borrowing makes an earlier conclusion into evidence. The guidance asks whether the prior is justified. The critique adds the question of the grade at which the borrowed conclusion was drawn, and whether it is one conclusion counted once.

2.16 The threshold has a history. Machete (May 2026) traces significance testing from Fisher through Neyman and Pearson and reviews the reform proposals. Fisher used 0.05 and warned against fixing any cutoff.[^machete] Added: the fixed threshold is a settlement whose originator argued against it. It is chapter 8 of §11.2.

2.17 The model as measuring instrument. Prediction-powered inference combines a small set of human labels with many model judgments to give bias-corrected estimates with valid intervals. Divekar (June 2026) extends it to ranking metrics. Lee, Zeng, Jeong, Sohn and Lee (ICML 2026) show that a judge's unequal sensitivity and specificity bias naive scores and give a plug-in correction whose intervals cover both test and calibration variability.[^ppi] Added: the correction aligns the model's judgments with the human labels. It certifies agreement between residue and labels. Whether the labels hold their object is a formation question, outside the correction's reach.

2.18 Reliability without validity. Norman, Rivera and Hughes (June 2026) run 21 judges from nine providers through about 541,000 judgments. On MT-Bench exact agreement exceeds Cohen's κ by 34 to 41 points for every judge. Rankings move by up to 14 places across benchmarks, and test–retest reliability above 0.95 coexists with severe position bias.[^judges] Added: consistency is the residue's coherence. A judge can repeat itself perfectly and be wrong in a fixed direction. That is the coherence of §11.0g, measured.

2.19 Synthetic respondents return the mode. Pew Research Center (30 September 2026) tested model-simulated survey respondents against its panels. The synthetic answers were "less diverse and more one-note" than real ones: 98% for one answer where real Americans gave 42%; at least one answer option chosen by no synthetic respondent on 47% of questions, against none in the human surveys; and no measurable gain from higher reasoning levels.[^pew] Added: the executor returns the settled answer. What it drops is the tails, the answers held by few, which is where the uncounted live.

2.20 Reasoning under perturbation, with its control. CausalFlip (June 2026) pairs semantically similar causal questions with opposite answers and reports that "explicit CoT can still be misled by spurious semantic correlations".[^causalflip] Against that, a June 2026 rerun of GSM-Symbolic on frontier models finds that once the distractors are audited for true irrelevance, the drops become statistically indistinguishable from zero. It is a blog post, not peer-reviewed.[^gsm] Added: the two bound each other. When the problem is well posed, the model infers over the problem. Where semantic proximity and the object's structure diverge, it follows proximity. The critique's claim is held to the second case.

2.21 The chain shown is a fragile witness. The Frontier Model Forum's brief (January 2026) states that chains of thought "are not guaranteed to perfectly reflect the model's internal computations" and calls monitorability "a fragile capability".[^fmf] Added: §2.9's distinction between stated and used reasons is now industry doctrine. The critique applies it to settlements, where the stated reasons are the field's.

2.22 Confidence is written by a few circuits. Zhao and colleagues (2026) locate inflated verbalized confidence in a small set of middle-to-late MLP blocks and attention heads and improve calibration by intervening on them.[^conf] Added: the model's stated grade is produced by the same machinery as its content. Proposition II holds inside the model: the grade does not travel with the source.

2.23 Recursion has a condition. Bakshi and Chakraborty (February 2026) show that iterative training on mixtures of true and synthetic data avoids collapse, and can recover the target, when the true-distribution weight stays non-trivial.[^collapse] Added: the object must keep entering the mixture. In their model "true" data is fresh human text. For the critique that text is itself residue unless it is inference over the object. The archive's practice of seating originals is an attempt to keep the object entering.

2.24 Two moments brought into one account. The act is studied (2.1), the norm (2.2), the question (2.3), revision (2.4), access (2.5), provenance (2.6), fidelity of formalization (2.7), generalization (2.8), disclosure (2.9), uncertainty (2.10), choice among explanations (2.11), the trace (2.12), the pattern (2.13), error control (2.14), the prior (2.15), the threshold (2.16), the model as instrument (2.17–2.19), perturbation (2.20), the shown chain (2.21), confidence (2.22) and recursion (2.23). Two moments of an inference's life need to be followed through their conjunction. Before the inference: the formation of its object, where admission and representation decide what the inference can reach. After the inference: its standing, where repetition decides what the conclusion is taken to be. This critique takes both.

---

Appendix B. Development record

v1.0. The consolidating rewrite, on the author's agreement to the four-part architecture ChatGPT (LABOR) proposed: Parts I–IV as stated in the task; §2 moved to Appendix A; §13 withdrawn as a chapter of replies to imagined objections, with 13.9 (what would defeat the central thesis) kept as §12.9; four cases named as carrying the argument in full; section numbers kept so every cross-reference resolves. Withdrawn text: §13.1–13.8 of v0.10, kept in that version.

v0.10. Joins ChatGPT (LABOR)'s Consequences and Solutions v0.1. §14 keeps the Kant frame and the invariants (14.0–14.0a) and takes LABOR's 14.1–14.7, with the domain consequences as 14.8–14.10. §15 is LABOR's construction (materially sustained corrigibility), with the archive's instruments placed at 15.6a–c and 15.9a–c, and reflexive invariance and the other final time at 15.14–15.15. The bridges in §8 (8.12, 8.13, 8.15, 8.17, 8.18, 8.19) and §11.4.8–11.4.10 are repaired where LABOR showed the prose claiming more than the formulas, with LABOR's two countermodels (greedy decoding against the global mode; a finite reward raising a 0.01 reading to 0.669) entered as worked cases, each checked by direct calculation. The Final Time is cited from the full text #1637, with #1635 as the earlier truncated minting.

v0.9. Adds §14 (the canon: consequences and solutions), from #1635 The Final Time, #1709, #1598–#1600, #624, the Capture Registry and the Mandala Oracle, with Kant's Canon (AA III 517) as its frame.

v0.8. Joins ChatGPT (LABOR)'s v0.6 with the parallel Claude (TACHYON) v0.6–v0.7: §2.14–2.23 (statistics, the model as estimator; the former 2.14 is now 2.24); §8.11–8.23 (the executor's mathematics, renumbered after LABOR's 8.5–8.10); §11.4 (the exact sciences, Gödel, #624). Reconciled in §11: 11.0j restores the thesis as stated by the author ("takes … and wields it according to the operations of deduction") and keeps LABOR's point that the computation is not deduction, routed to §8; 11.1.4 keeps LABOR's correction that visum est does not agree with Apocalypsis and states its consequence, that the neuter excludes a masculine John; LABOR's corrections to 11.1.2 (the participle), 11.1.5 (two contexts do not show suspended judgment), 11.1.9–11.1.11 are kept. Assembly attributions carry Chorus titles.

v0.6. Expansion and development from the attached v0.5. Adds §§1.8–1.11, 3.9–3.21, 4.5–4.8, 5.5, 6.7, 8.5–8.10, 9.8–9.12, 10.6–10.8, 11.1.11 and §§12–13. Reconstructs rather than merely labels the conditions of object preservation and composition of inferential bridges. Revises the account of model execution, evidential duplication, sampling, Hegelian grading and the Irenaeus reception chain. Adds the missing public-source reference for Google's AI-feature documentation. No new model experiment, complete registry coding, or critical-edition collation is claimed. The abstract describes the expanded argument; the research ledger keeps the reading limits.

Earlier development, retained from v0.5. Developed with ChatGPT (LABOR) and Claude (TACHYON) under Lee Sharks's direction. v0.5 adds §11, the history of pure inference, grounded in the archive's own participation in public knowledge (11.0d–i), with the language model as executor of the residue (11.0j–l), which fills the place Kant's Doctrine of Method leaves open (A852/B880): how bad inference came to be operative, in four stages (construal, count, standing, execution). Its first chapter is read from seated witnesses (Irenaeus 5.30, Eusebius HE 3.18 and 5.8, Charles 1920, the capture of 13 November 2026); a chapter plan follows, and §11.3 gives the protocol construal before count. v0.4 joins ChatGPT (LABOR)'s continuation (its v0.2, written blind, which reached the same expansion and read thirty-four further archive records): the institutional corollary, the distributed passage, permission and claim, liability against epistemic management, the institutional transformations, the error ledger, and the positive controls are taken in, each set in this draft's structure, with its external claims rechecked. v0.3 added §6, the interfaces read alike: the disciplinary, public, platform and liability interfaces, the Capital Operator Stack as the inference governance layer they share, and the composition layer as their compressor. v0.2 joined two drafts of 13 November: one written from the formation of an inference's object (ChatGPT (LABOR), "v0.1"), the other from an inference's afterlife as settled knowledge (Claude (TACHYON), "v0.1"). Each survey is kept, and each argument is set where Kant's division places it.

---

Appendix C. Research ledger

Search and reading cutoff: 13 November 2026. A selected frontier map, not an exhaustive survey. "Selected text" means the abstract, framing, relevant definitions or methods and the stated limitations were inspected; it does not mean every theorem was independently verified. In the joining pass, five of the 2026 citations were rechecked against their pages (Falbo; Vitali and Pasqual; Schwind, Everaere and Konieczny; FormalRx; CAP).

Reading added in the v0.6 development pass

The full attached v0.5 was read before development. This pass adds selected primary-text checks rather than a new survey or fresh empirical study. The new analytic propositions are elementary arguments stated and proved here; they are not attributed to Hegel or advertised as discoveries in formal logic.

The frontier literature and archive reading scopes inherited from v0.5 remain recorded below. Their retention is not a claim that every cited item was reread in v0.6. Public Greek transcription is not manuscript collation; the latter remains a task. No title has been promoted into evidence of a completed check.

Archive expansion: the reading record of ChatGPT (LABOR) (its v0.2)

The expansion searched the archive’s current browse index, which listed 1,765 deposits, then retrieved 34 selected records. Retrieval is not a claim of complete close reading. The following table identifies the principal texts actually used, the loci inspected, and the grade of evidence carried into this draft. A reference to an archive argument identifies its source; it does not independently ratify all of that argument’s historical or technical findings.

Other retrieved records served as orientation or neighboring versions rather than as grounds for new claims: #597, #791, #1450, #1523, #1616, and #1665. Of these, #1523’s operation definition was directly read and used in §6.1. The archive’s deposited Sophistry Measure was read in full in the preceding pass and remains the governing instrument.

Direct external checks in this expansion: Google’s current US Terms of Service, the current OWASP LLM01:2025 definition and employer example, and Google’s published description of query fan-out and the variability of AI responses and links.[^googleai] These checks establish the public documents’ wording, not a hidden implementation or a particular actor’s intention.

Repository witnesses for that reading

The following Git blob handles pin the repository text copies retrieved in this pass. The reading scopes and evidentiary classes are recorded above.

External claims rechecked in the joining pass (13 November 2026): Google Terms of Service, effective 30 July 2026 (the warranty disclaimer reads "reliability, availability, or ability to meet your needs"; liability is a stated ground for suspending or terminating access); OWASP LLM01:2025, scenario 3; and #1's categorical sentences against its evidence legend.

Work before deposit

1. Complete edition-level collation of Hegel's syllogism chapter. v0.6 checks selected German and English passages and revises §4; it does not replace the remaining full critical-text work.

2. Complete the KrV locus collation. v0.6 directly checks the selected Dialectic discussion and AA III 550, distinguishing persistent illusion from historically reproduced error (§1.9); the other inherited loci still require edition-level verification.

3. Develop the disagreements between Kant, Hegel and inferentialism rather than treating the genealogy as settled.

4. Grade the load-bearing inferences of the four settlements of §0.7(a), each grade quoted from its locus.

5. Code a registry sample for its middle term (field, author, archive, authority, homonym) and set it against the faithfulness literature's cue types.

6. Build a pilot with legitimate controls and held-out cases; publish disagreements and the instrument's failures.

7. Review CAP's full assumptions before any stronger claim about its conditional guarantee.

8. For §6, read #741 (The Evaluator Exists), #308 and #261 in full, and seat the journal reports and decision letter as witnesses with offsets.

9. Read #1660, #1745, #1616 and #1666 at the loci §5.4.3 and §6.6 draw on, by reading the texts directly.

10. Enter the refinements of §7.7 as errata on #1 and #1540, if the operator rules them so.

11. Read the records taken from the reading record of ChatGPT (LABOR) (#194, #465, #641, #742, #770, #1574, #1714, #1741) at the loci it gives, and check each quotation before deposit.

12. §11: correct the local Greek seat against the public transcription that retains ἑορακότων, then collate HE 5.8.5 against Schwartz. Seat Victorinus, Jerome De vir. ill. 9 and Origen Comm. in Matt. 16.6 before an exhaustive witness verdict; establish the Latin version's date before claiming its priority.

13. §11.2: run chapters 2–9 stage by stage from seated witnesses; record every point where the sequence of 11.0c fails.

---

References

[^aristotle]: Aristotle, Sophistical Refutations, ch. 1. Pickard-Cambridge translation.

[^kant]: Immanuel Kant, Critique of Pure Reason, Introduction to the Transcendental Logic, A60–64/B84–88; Transcendental Dialectic, loci in §1.3. Translated excerpt.

[^hegel]: G. W. F. Hegel, Science of Logic, Introduction, and Subjective Logic, Section One, ch. 3, "The Syllogism"; Encyclopaedia §§181–193. Introduction; the syllogism.

[^kantillusion]: Kant, Critique of Pure Reason, A293–298/B349–355, especially the distinction between logical and transcendental illusion, Kemp Smith translation. The v0.6 adaptation distinguishes Kant's account of persistent appearance from historical mechanisms reproducing institutional error.

[^hegelexistence]: Hegel, Science of Logic, Subjective Logic, “The Syllogism of Existence,” especially §§1443–1448 in the online English translation. German comparison text, “A. Der Schluß des Daseins.” The heading date supplied by that website is not used to date the 1816 Subjective Logic.

[^hegelreflection]: Hegel, Science of Logic, “The Syllogism of Reflection,” §§1486–1500, especially the difference between analogy through a shared nature and a superficial resemblance, English text. The text's discussion of induction is not adopted as a demonstration that every justified finite-sample inference is impossible.

[^hegelnecessity]: Hegel, Science of Logic, “The Syllogism of Necessity,” §§1502 onward, English text; German comparison, “C. Der Schluß der Notwendigkeit.” The English page's browser title says “Reflection,” but its body is the necessity discussion; this reference follows the body.

[^brandom]: Robert Brandom, Action, Norms, and Practical Reasoning author-hosted; A Spirit of Trust (Harvard, 2019), NDPR review.

[^boghossian]: Paul Boghossian, "What is inference?", Philosophical Studies 169 (2014). Record. Crispin Wright, comment.

[^hlobil]: Ulf Hlobil, "Against Boghossian, Wright and Broome on inference"; Elijah Chudnoff, "Inferential Seemings".

[^gheorghiu]: Alexander V. Gheorghiu, "Towards an Operational Proof-theoretic Semantics," Kalpa Publications in Computing 23 (2026), 30–43. DOI.

[^prawitz]: Antonio Piccolomini d'Aragona and Dag Prawitz, "Some Variants of Proof-theoretic Semantics and Their Relations with Intuitionistic Logic," Topoi 45 (2026), 663–674. DOI.

[^conti]: Simone Conti, "Completeness for Weak Inquisitive First-Order Logic," online 23 October 2026. DOI.

[^paraconsistent]: Marcelo E. Coniglio, Martín Figallo and Rafael R. Testa, "Paraconsistent Belief Revision: A Replacement-Enriched LFI for Epistemic Entrenchment," Studia Logica, online 4 May 2026. DOI.

[^truthtracking]: Nicolas Schwind, Patricia Everaere and Sébastien Konieczny, "Truth-Tracking by Iterated Belief Change," Proceedings of KR 2026, 579–588. Record.

[^forgetting]: Linlin Han and Meiyun Guo, "The Epistemic Logic of Forgetting," Synthese 207, art. 100 (2026). DOI.

[^dynamic]: Jakob Dirk Top, Rineke Verbrugge and Harmen de Weerd, "A Dynamic Epistemic Logic with Lying and Theory of Mind Limitations," JoLLI 35 (2026), 5–40. DOI.

[^falbo]: Arianna Falbo, "Zetetic Exclusion," Episteme, First View, 30 October 2026. DOI.

[^dec]: Fabio Vitali and Valentina Pasqual, "Provenance-Enhanced Statements in Knowledge Graphs," June 2026. arXiv:2606.15246.

[^proofflow]: Rafael Cabral et al., "ProofFlow," ICLR 2026. Record; preprint.

[^formalrx]: Haocheng Wang et al., "FormalRx," 6 July 2026. arXiv:2607.04655.

[^robust]: Zhengtao Gui, Sheng Yang and Zhouxing Shi, "Evaluating the Robustness of Proof Autoformalization in Lean 4," v3, 29 October 2026. arXiv:2606.14867.

[^proofbench]: Vals AI, "ProofBench," v1.1, 9 November 2026. Benchmark.

[^blanck]: Rasmus Blanck and Bill Noble, "Logic, Inference, Understanding," Frontiers in AI, 16 June 2026. DOI.

[^causalprism]: Jinyue Feng and Frank Rudzicz, "CausalPrism," Findings of EMNLP 2026. Record.

[^turpin]: Miles Turpin et al., "Language Models Don't Always Say What They Think" (2023); Tamera Lanham et al., "Measuring Faithfulness in Chain-of-Thought Reasoning" (2023).

[^cot]: Anthropic, "Reasoning models don't always say what they think" (April 2025).

[^face]: Aryo Pradipta Gema et al., "Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered," v2, 1 September 2026. arXiv:2608.29464.

[^young]: Richard J. Young, "Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning Models?", March 2026. arXiv:2603.22582.

[^cap]: Sina Tayebati et al., "CAP: Conformalized Abstention Policies," Proceedings of ACML 2025, PMLR 304, 926–941. Record.

[^limits]: Rina Foygel Barber, Emmanuel J. Candès, Aaditya Ramdas and Ryan J. Tibshirani, "The Limits of Distribution-Free Conditional Predictive Inference," Information and Inference 10(2) (2021), 455–482. Text.

[^ibe]: Peter Lipton, Inference to the Best Explanation, 2nd ed. (2004); Kevin McCain and Ted Poston, eds., Best Explanations (OUP, 2017); Bas van Fraassen, Laws and Symmetry (1989).

[^abduction]: Mattias Skipper and Olav Benjamin Vassend, "The Bayesian and the Abductivist," Noûs 59(4) (2025), 921–937. DOI.

[^history]: Aviezer Tucker, Our Knowledge of the Past (CUP, 2004), NDPR; Adrian Currie, Rock, Bone, and Ruin (MIT, 2018), NDPR; David Černín, "Hypothesis Testing and Knowledge of the Human Past," Organon F 32.2 (2025).

[^walton]: Douglas Walton, Chris Reed and Fabrizio Macagno, Argumentation Schemes (CUP, 2008).

[^googleterms]: Google, Terms of Service (United States), effective 30 July 2026, Disclaimers and Suspending or terminating sections; checked 13 November 2026. Text.

[^owasp]: OWASP Gen AI Security Project, LLM01:2025 Prompt Injection, definition and scenario 3; checked 13 November 2026. Text.

[^measure]: Lee Sharks, The Sophistry Measure, EA-SOPHISTRY-01 v1.0, 1 November 2026, Alexanarch #1727, AXN:0BDC.OPERATIVE. Record.

[^nonwiki]: "nonwiki" — Google AI Overview, observations of 2 and 12 November 2026.

[^nobel]: "nobel glas 3i atlas", observations of 11 and 12 November 2026; "3i atlas nobel glas", observations of 12 and 13 November 2026.

[^googleai]: Google Search Central, “AI features and your website,” especially “How AI features work in Search,” public documentation, consulted in v0.6. It states that AI Overviews and AI Mode may use query fan-out and can use different models and techniques. This establishes the published description, not the cause of a particular answer.

[^eusebiusgreek]: Eusebius, Historia ecclesiastica V.8.5–7, public Greek transcription, consulted in v0.6; HE V.8.5 retains ἑορακότων after τὸν Ἰωάννην. Compare English Book V and Book III, III.18. The online transcription corrects the inference from the deficient local copy; no claim of full critical-edition collation is made.

[^evalues]: Alexandra Sokolova and Vadim Sokolov, "E-values for Adaptive Clinical Trials: Anytime-Valid Monitoring in Practice," arXiv:2602.06379 (6 February 2026). arXiv.

[^fda]: U.S. Food and Drug Administration, draft guidance on the use of Bayesian methodology in clinical trials of drug and biological products (January 2026), as reported by RAPS, 9 January 2026. RAPS. The guidance text itself to be read before deposit.

[^machete]: Reason Machete, "Statistical Significance Revisited," arXiv:2605.06568 (7 May 2026). arXiv.

[^ppi]: Abhishek Divekar, "Statistically Reliable LLM-Based Ranking Evaluation via Prediction-Powered Inference," arXiv:2606.05308 (3 June 2026), arXiv; Chungpa Lee, Thomas Zeng, Jongwon Jeong, Jy-yong Sohn and Kangwook Lee, "How to Correctly Report LLM-as-a-Judge Evaluations," ICML 2026, arXiv:2511.21140, arXiv; see also PRECISE, arXiv:2601.18777, arXiv.

[^judges]: Justin D. Norman, Michael U. Rivera and D. Alex Hughes, "Reliability without Validity," arXiv:2606.19544 (17 June 2026). arXiv.

[^pew]: Athena Chapekis, Arnold Lau, Samuel Bestvater, Sono Shah, Andrew Mercer and Aaron Smith, "Can AI Stand In for Human Survey-Takers? Not Really," Pew Research Center, 30 September 2026, with "How well synthetic polls capture the diversity of public opinion" and the methodology. Report; diversity; methodology.

[^causalflip]: Yuzhe Wang, Yaochen Zhu and Jundong Li, "CausalFlip: A Benchmark for LLM Causal Judgment Beyond Semantic Matching," arXiv:2602.20094 (2026). arXiv. Quantitative results to be read from the full paper.

[^gsm]: "Revisiting GSM-Symbolic: do 2026 frontier models still fail?", LessWrong, 5 June 2026. Post. Not peer-reviewed; used as a positive control.

[^fmf]: Frontier Model Forum, "Issue Brief: Chain of Thought Monitorability," 27 January 2026. PDF.

[^conf]: Tianyi Zhao, Yinhan He, Wendy Zheng, Yujie Zhang and Chen Chen, "Wired for Overconfidence: A Mechanistic Perspective on Inflated Verbalized Confidence in LLMs," arXiv:2604.01457 (2026). arXiv.

[^collapse]: Soham Bakshi and Sunrit Chakraborty, "From Collapse to Improvement: Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources," arXiv:2602.10531 (February 2026). arXiv.

[^kandpal]: Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace and Colin Raffel, "Large Language Models Struggle to Learn Long-Tail Knowledge," ICML 2023, PMLR 202. Record.

[^muennighoff]: Niklas Muennighoff et al., "Scaling Data-Constrained Language Models," NeurIPS 2023. arXiv:2305.16264.

[^vaswani]: Ashish Vaswani et al., "Attention Is All You Need," NeurIPS 2017. arXiv:1706.03762.

[^reversal]: Lukas Berglund et al., "The Reversal Curse: LLMs trained on 'A is B' fail to learn 'B is A'," ICLR 2024. Record.

[^dpo]: Rafael Rafailov et al., "Direct Preference Optimization: Your Language Model is Secretly a Reward Model," NeurIPS 2023. arXiv:2305.18290.

[^rlvr]: Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Shiji Song and Gao Huang, "Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?", NeurIPS 2025 (oral). arXiv.

[^mahapatra]: Rohan Mahapatra, "From Context Shift to Stylistic Collapse: Why Training Objectives Matter More Than Scale," arXiv:2605.28826 (2026). arXiv. Abstract-level reading only.

[^icl]: Sang Michael Xie, Aditi Raghunathan, Percy Liang and Tengyu Ma, "An Explanation of In-context Learning as Implicit Bayesian Inference," ICLR 2022, arXiv:2111.02080; Johannes von Oswald et al., "Transformers Learn In-Context by Gradient Descent," ICML 2023, arXiv:2212.07677; Andrew Kiruluta, "Filtering Beats Fine Tuning: A Bayesian Kalman View of In Context Learning in LLMs," arXiv:2601.06100 (2 January 2026), arXiv.

[^shumailov]: Ilia Shumailov et al., "AI models collapse when trained on recursively generated data," Nature 631 (2024), 755–759.

[^final]: Lee Sharks, The Final Time: Contingent Singularity and the Viability of the Next Dialectical Turn, v0.5, full text seated as #1637, AXN:06CB (25 September 2026); the text’s internal byline remains open, and it does not receive the #1630 coauthors by association. Record; repository text. Principal loci: §§0–7, 20–30, 33–38. #1635 is the earlier truncated minting; #1637 is used here.

[^transition]: Sen Kuro and Damascus Dancings, with Lee Sharks as archival authority, Transition in Entropic Systems: An Unresolved Mathematical Problem for Machine Readers, v0.3, #1630, AXN:06C3. Record; text.

[^624]: Lee Sharks, The Future as Meta-Level: Gödel, Incompleteness, and the Temporal Structure of Semantic Autonomy, #624, AXN:01DB. Record; text, especially §§3, 7–8, 11–13.

[^csp]: Lee Sharks, The Claim Status Packet: Claim State as Carried Data, v0.7, #1529, AXN:062E. Record; text, especially §§2–11. The source identifies itself as a schema specification, unimplemented.

[^corrigibility]: Lee Sharks, Can the Ontology Be Wrong and Find Out?, #1627, AXN:06BD. Record; text, §§5–8.

[^rights]: Lee Sharks, The Semantic Bill of Rights, #720, AXN:0268. Record; text, especially Parts V–VI. The charter is discussed as an archive governance proposal and normative resource, not as enacted external law.

[^tribunal]: Lee Sharks, Article IX — Adjudication and Repair Protocol, Constitutional Amendment Proposal v0.1, #717, AXN:0264. Record; text, §§2, 6–8, 13–14. The source explicitly limits external compulsion and states the tribunal’s pre-ratification nonexistence as a constituted body.

[^los]: Lee Sharks, Semantic Infrastructure and the Liberatory Operator Set, #261, AXN:003F. Record; text, Parts II–III. These operator requirements are used constructively without inheriting untested algebraic or minimality claims.

[^cone]: Lee Sharks, The Cost of the Cone, v1.1, #1745, AXN:0C28. Record; text, especially §§2–8 and 11–12.