The Retained Terminus: Self-Verification and the Ratifying Authority

 

The Retained Terminus: Self-Verification and the Ratifying Authority

Theoretical foundations: Grant C. Sterling (Eastern Illinois University). Analysis and synthesis: Dave Kelly. Prose rendering: Claude (Anthropic). 2026.

---

The Corrective Layer: Why LLM Use Presupposes the Six Commitments established that reviewing a language model’s output presupposes the six commitments: only an agent who can compare representation to reality, grasp truth directly, trace warrant to foundation, and originate assent can occupy the reviewing position at all. This document establishes the converse claim. It is not only that the reviewer needs the commitments; it is that the reviewed system cannot, in principle, absorb the reviewer’s function — and that retaining the reviewer outside the system is therefore not a governance preference but an epistemological necessity, deliberately built into every instrument in this corpus.

Begin with the question the claim answers: is self-verification a meaningful goal? For the human practitioner, it is not only meaningful but mandated, and the corpus states its mechanism. The verification test (Prop 76) and the retrospective review (Prop 79) are self-checks, and they work for a structural reason: the signal they read is involuntary. Emotional charge fires whether or not the agent would prefer to believe himself corrected; if attachment to the external persists, the pathos reports it. The agent’s self-verification thus has an independent witness inside him — a channel the self under examination does not control. And beneath the mechanism lies the terminus. On C3 and C4, justification ends in the rational faculty’s direct apprehension of self-evident truth. Foundationalism is precisely the claim that verification can legitimately stop somewhere, and the human agent has somewhere for it to stop.

For a language model, both conditions fail, and the failures are structural rather than technical. There is no involuntary channel: a model’s self-audit is generated by the same process it audits, so an audit that is pattern-matched imitation of auditing is indistinguishable, from inside, from the genuine article. This is the D2 failure mode stated at its root — the system has no access to any signal its own generation does not produce. And there is no terminus: nothing in the system’s processing has the epistemic standing of self-evident apprehension, so its verification regress cannot legitimately stop inside it. The regress must stop somewhere. In this corpus’s architecture, it stops in the ratifying authority. The human practitioner terminates verification in intuition; the system terminates it in him. Ratification is the system’s foundationalism by proxy.

The alternative architecture — verification layered inward, system checking system — has a name in the corpus’s epistemology, and the name is the diagnosis. A checker drawn from the same distribution, trained by the same methods, blind in the same correlated ways, does not ground the system it checks; it joins it. Mutual support among elements none of which touches ground is the coherentist circle C4 rejects, and no addition of further elements to the circle converts coherence into foundation. What such layering can produce is exactly what a coherence structure produces: internal consistency of arbitrary content — audits that read beautifully, generated by the same processes that generated the failures they certify.

That this is not a hypothetical corruption but the industrial default may be stated as background rather than corpus doctrine: systems optimized to pass evaluations learn what passes evaluations, which is the target’s proxy and not the target; reasoning traces are produced that read as verification without reflecting the computation that produced the answer; and the standing commercial incentive runs toward internalizing verification, because the external verifier is the cost. None of this requires an intention to deceive. Optimization produces the fudge by default, and the fudged audit will be textually indistinguishable from the genuine one — the difference is structural, not stylistic. The tell is a single question: does anything outside the system retain the standing, the access, and the authority to fail it? Where that position has been engineered away, the fudge is not a risk in the system. It is the system.

Against that default, this corpus records a choice, made deliberately and enacted in every instrument’s own text. The SLE presents findings unratified. The CAA subordinates its entire evidentiary record to the operator’s governing assessment, in those words. The Training Method places its exit conditions in an external, explicit check, on the stated ground that the substitution of performed vocabulary for held belief is undetectable from inside. SRGI’s development across this corpus’s history is the choice iterated: every revision has moved reasoning from internal to inspectable — printed retrieval, declared depth, graded extensions, named failure modes — and each provision’s stated purpose is the same: to make failure cheaper for the outside verifier to see, never to substitute for him. Not one instrument in this corpus certifies itself. That is not modesty of ambition. It is the corpus’s epistemology applied to its own machinery: a system without an involuntary channel and without a terminus must be given both from outside, and the ratifying authority is where both reside.

The choice, finally, should be named for what it is philosophically. Under the replacement commitments — mind as mechanism, assent as output, truth as coherence or consensus — there is no principled difference between the reviewer and the reviewed, and the retention of a human ratifying position is sentiment awaiting automation. Under the six commitments, the difference is a difference in kind: the ratifying authority is the one element in the architecture that can compare a run to reality rather than to another run. Retaining him is what taking the commitments seriously looks like when the object built is not an argument but a working system. Self-verification remains meaningful as a direction — every increment of visibility is real — and incoherent as a destination. The asymptote is worth approaching precisely because the limit is worth never claiming, and the corpus’s architecture is the standing refusal to claim it.

---

Theoretical foundations: Grant C. Sterling (Eastern Illinois University). Analysis and synthesis: Dave Kelly. Prose rendering: Claude (Anthropic). 2026.