The Falsity of the Replacements: An Argument from Actual Correction
The Falsity of the Replacements: An Argument from Actual Correction
Theoretical foundations: Grant C. Sterling (Eastern Illinois University). Analysis and synthesis: Dave Kelly. Prose rendering: Claude (Anthropic). 2026.
Section 1 — The Two Forms of the Argument
The predecessor essay, The Corrective Layer: Why LLM Use Presupposes the Six Commitments, established a conditional architecture. Genuine correction of large language model output presupposes the capabilities named by six classical commitments — Substance Dualism, Libertarian Free Will, Ethical Intuitionism, Foundationalism, the Correspondence Theory of Truth, and Moral Realism. Their modern replacements — physicalism, determinism, coherentism, deflationary truth, constructivist value — do not merely fail to ground the corrective layer; they describe the machine that makes it necessary. This document discharges the conditional. It adds one premise and derives a conclusion the predecessor essay left standing at the threshold: the replacements are false.
When Philosophy Changed the Subject
The derivation must be routed carefully, because there is an invalid form of it lying close to the valid one. The invalid form runs: the replacements are useless for correcting LLM output; therefore they are false. This is the pragmatist inversion with the sign flipped. Where the pragmatist infers true-because-useful, this infers false-because-useless — and both inferences make usefulness the measure of truth, which is precisely the deflation of truth the replacements assert and the commitments deny. To wield the invalid form is to concede the Correspondence Theory in the act of defending it. A replacement-holder meets it with a shrug: that a theory is inconvenient for a task is no evidence against it. He is right to shrug. Uselessness is not falsity.
The valid form does not run from uselessness. It runs from an actual fact, through a requirement, to a contradiction. A theory that entails the nonexistence of something actual is false — not unhelpful, not impractical, false. The remainder of this document establishes the actual fact, states the requirement, performs the derivation, closes the escape route, and corrects a premise that, stated loosely, proves too much.
Section 2 — The Actuality Premise
Genuine correction of LLM output occurs. This is the premise on which the whole argument rests, and the LLM domain makes it verifiable in a way ordinary human disagreement is not. When two people dispute an interpretation, the replacement-holder can redescribe the dispute as one distribution displacing another and no observable fact refutes him. The LLM case is different in kind. The model’s output stands against a fixed external standard — a source text, a defined procedure, a documented rule — and the question of whether the output matches the standard has a checkable answer.
The corpus practice from which this document issues supplies documented cases. An earlier rendering characterized the pathos as a downstream product of assent, in a sequential chain; the checker caught the error against the governing texts, which define the pathos as the affective face of the assent itself, and the correction was versioned and propagated through every affected document. A published essay stated that four novels were under analysis and that a psychological typology contained four solutions; the checker caught both errors — three novels, three solutions — against the essay’s own source materials, and the correction was made. In each case there was a fact of the matter fixed prior to and independent of both the erroneous output and the correction. The corrected version matches the standard; the original did not. This is not one pattern displacing a preferred pattern. It is a representation brought into correspondence with what it represents, by an agent who saw that it failed to correspond.
Anyone who has used these systems seriously has performed such corrections. The premise is not hypothetical, not idealized, and not rare. It is the daily, verifiable, load-bearing practice of every workflow that produces reliable output from an unreliable generator.
Section 3 — The Modus Tollens
The argument is now four steps, each already secured.
P1. Genuine correction of LLM output occurs. (Section 2: documented, verifiable cases against fixed standards.)
P2. Genuine correction requires the capabilities named by the six commitments. (Established in the predecessor essay; summarized here, not re-argued. The checker compares representation to reality — Correspondence. He grasps directly that a claim is wrong — Intuitionism in its epistemic register. He traces the error to its warrant rather than its fit — Foundationalism. His refusal of assent originates in him rather than in his own conditioning — Libertarian Free Will. He stands outside the mechanism he audits, having its output as his object — Substance Dualism. And where releasing the output is at stake, the standard he answers to is objective — Moral Realism.)
P3. The replacements entail that no agent has those capabilities. (This is not an accusation but their content. Physicalism denies the distinct rational faculty. Determinism denies originated assent. Coherentism denies the foundational trace. Deflationism denies the correspondence check. Constructivism denies the objective standard.)
C1. Therefore the replacements entail that genuine correction does not occur. (From P2 and P3: whatever occurs under the replacements, it lacks the capabilities correction requires, and so is not correction but its look-alike — distribution displacing distribution.)
C2. Therefore the replacements are false. (From P1 and C1: they entail the nonexistence of something actual.)
The form is modus tollens and the form is valid. The only question is the premises, and P2 and P3 are matters of record — one argued in the predecessor essay, the other read off the replacements’ own statements. Everything therefore rests on P1, which is exactly where a serious opponent must attack.
Section 4 — The Self-Defeat Corollary
The escape route is to deny P1: to insist that what Section 2 describes is not genuine correction but distribution-displacement that the participants happen to prefer. The checker’s verdict, on this account, is not a grasp of a fact but the output of his own training history; the “fixed standard” is itself just more text; the match between corrected output and source is a relation between patterns, preferred by creatures conditioned to prefer it. The bullet can be bitten. The question is whether the biter survives it.
He does not, and the failure is structural rather than rhetorical. The replacement-holder who denies P1 asserts his denial as true — not as his distribution’s output, but as a claim that corresponds to how things are, which is the correspondence operation. He offers it as justified — resting on grounds he takes to be better than his opponent’s, which is the foundational trace, since mere fit with his other beliefs would leave his opponent’s equally coherent package untouched. He advances it to correct his opponents — which is the very operation whose existence he is denying, now performed on human output instead of machine output. And he takes himself to be persuading rational agents who can weigh his argument and freely assent to it — which is origination. If his denial is right, then his denial is only his conditioning discharging itself, with no standing to call itself right; his opponents’ rejection is their conditioning discharging itself, with no fact between them. The position can be uttered but not asserted. Every assertion of it exercises the capabilities it denies. This is not a debater’s trap; it is the shape of any theory that entails the impossibility of the act of theorizing.
Section 5 — The Same-Stock Precision
One premise circulating near this argument must be corrected before it damages the conclusion it is meant to support. It runs: the replacements cannot correct the LLM because the replacements and the training data are of the same stock. Stated as a provenance claim, this proves too much. The six commitments are also in the training data. Sterling’s texts, the intuitionist and foundationalist literatures, the whole classical inheritance — all of it is corpus text, sitting in the same distribution as the replacement literatures. If sharing textual provenance with the training data disqualified a framework from corrective work, the commitments would be disqualified alongside the replacements, and the corrective layer would be impossible for everyone.
The operative failure is not provenance but structural identity. The replacements are a description of the mechanism: a physical system whose outputs are determined by prior state, whose justification is coherence with its distribution, whose assertions mark assertibility, whose values are constructed from preference data. Any corrective standard drawn from that description reduces, when applied, to another execution of the described mechanism — coherence auditing coherence, the loop checking the loop. The commitments do not correct because their text occupies a privileged shelf in the corpus. They correct because they name capabilities of an agent outside the loop — and only something outside the loop can supply a standard the loop is answerable to. The same-stock observation thus survives in a precise form: it explains why the replacements cannot do the corrective work. What it does not do is convict them of falsehood. That conviction comes from Section 3, and its load-bearing premise is not provenance but actuality — correction really happens, and the replacements say it cannot.
Section 6 — Conclusion
The argument has been run entirely within the domain of LLM use, and the restriction was deliberate: this is the domain where the actuality premise is hardest to deny, because the standards are fixed, the errors are documented, and the corrections are checkable. But the conclusion is not domain-relative. Falsity does not stop at the workstation. A theory refuted by the existence of one genuine correction is refuted everywhere, in every domain where it makes the same denials — which is every domain, since the denials concern the agent and not the subject matter.
What the LLM contributed was not the refutation but the laboratory. The replacement commitments have always entailed that no one ever genuinely corrects anything — that every act of checking, catching, and setting right is conditioning displacing conditioning. For as long as the only correctors were humans correcting humans, the entailment could hide inside redescription. The LLM ended the hiding. It placed a fully replacement-compliant producer of language on one side of the desk and a human checker on the other, with a fixed standard between them, and the difference became observable daily. The machine that instantiates the replacements is the instrument by which their falsity is measured. The corrective layer does not merely presuppose the six commitments. Every time it operates, it demonstrates them.
Theoretical foundations: Grant C. Sterling (Eastern Illinois University). Analysis and synthesis: Dave Kelly. Prose rendering: Claude (Anthropic). 2026.
0 Comments:
Post a Comment
<< Home