Foundational thesis

Conditions for Ending Value Inquiry

Stopping is not justified merely because an agent feels certain. For the normative question at issue, genuine competing possibilities must disappear not only relative to the agent's present capacities, but also relative to cognitive expansions that are reasonably reachable from its epistemic and technological position.

Working thesis v0.6 · English translation · 2026-09-05

Thesis: Terminal closure of value inquiry should require more than high probability, strong conviction, valence, or reward signals. Two current candidate routes are: (A) normative force is presented at a foundational layer, and that presentation carries epistemic warrant for the target proposition without treating foundational status itself as sufficient warrant; that warrant must survive self-criticism and revalidation through the recursively reachable limit of cognitive expansion; or (B) the existence or nonexistence of the relevant normativity is logically closed strongly enough to eliminate remaining realist possibilities. Practical suspension while the question remains unresolved must be kept distinct from terminal closure.

Value-structure refinement: stopping is query-relative

Terminal closure is relative to the normative question being asked. Settling C does not by itself settle B or R, and foundational presentation may close only the particular question whose needed structure is presented rather than all of ethics.

Logical closure should likewise be distinguished into at least three forms: existence closure (whether any relevant normative structure exists), closure of a specific normative structure, and nonexistence closure. A stopping argument must say which of these it has established.

The demand to keep the hypothesis space open is not a demand to consider every logically describable fantasy. What matters are hypothesis-space expansions and conceptual advances that are reasonably reachable from the agent's epistemic situation and could bear on the question at issue.

1. Separate stopping conditions from inquiry allocation

This page asks not how many resources should be spent on inquiry, but when the target question may be treated as settled. An unresolved question may rationally receive little current inquiry or even none. Resource allocation is treated in Allocating Inquiry Under Unresolved Normative Uncertainty.

Thus unresolved ≠ continuous active inquiry, and not inquiring now ≠ resolved.

2. Generating a doubt is not the same as retaining a genuine rival

The clarity required for closure is not psychological certainty of 100%, nor an inability to emit the sentence “really?”. An agent can generate skeptical strings even about well-understood logical facts.

The issue is whether, after fully applying its present reasoning, self-criticism, and counterargument-generation capacities, and after adequately considering reasonably reachable cognitive expansions, the negative side remains an independent epistemic contender. If questions such as “is this merely desire?”, “is this only a reward signal?”, or “why is that a reason?” remain substantive live alternatives, then foundational normativity has not yet been presented to that agent in a stopping-condition form.

Clarity condition: What matters is not a feeling of being unable to doubt, but the inability—under present capacities together with reasonably reachable cognitive expansion—to construct an independent epistemic foothold for the opposing possibility.

“Reasonably reachable” does not mean every cognitively possible future one can imagine. It includes not only forms of cognitive expansion for which there is a realistic path from the agent's current knowledge, resources, technology, and capacity for self-modification, but also further expansions that become reachable only after those earlier expansions are achieved. A cognitive possibility is therefore not excluded from the stopping condition merely because the present agent cannot yet see a direct path to it.

Terminal closure requires that no genuine rival reopen even at the reachable cognitive limit obtained by iterating such reasonably reachable expansions. This does not require realizing every logically possible superintelligence or cognitive architecture; it is a condition over the cognitive space reachable through realistic paths from the agent's position.

3. Terminal route A: foundational presentation of normativity

One strongest candidate is a case in which value is given not merely as “I prefer this” or “this is unpleasant,” but together with its normative force.

However, being given at a foundational layer is not itself sufficient for justification. Route A requires both (i) non-inferential foundational presentation and (ii) justificatory force that gives epistemic warrant for the target proposition without deriving that warrant merely from the fact that the presentation is foundational. A foundational presentation may be an epistemic input not generated by inference, but it is not a “Given” exempt from inferential assessment. Whether it is genuine normative cognition must be tested within a web of inference that includes coherence with other beliefs, counterexamples, independent cognitive routes, self-audit, and increased cognitive capacity. A merely phenomenal or psychological given therefore does not satisfy Route A, however direct it may seem.

Route A therefore does not infer certainty from foundational status. It leaves open the possibility that a foundational presentation becomes certain cognition after surviving sufficiently strong validation.

Aversive pain, attractive pleasure, intense preference, or whole-person conviction is not enough if “why is that a reason?” still has substantive content. If instead asking that question eventually adds nothing beyond “why is something that is a reason a reason?”, and this status survives increased self-criticism and cognitive ability, the state may qualify as a candidate foundational presentation.

This page does not claim that humans or AI currently possess such cognition, or even that it is possible. It specifies a candidate condition for closure rather than claiming the condition is already met.

4. Terminal route B: logical closure from the inferential layer

Closure need not arise only through foundational presentation. It may also arise inferentially if the relevant possibility space is logically closed strongly enough.

Ordinary deduction or an inferential best explanation is not by itself sufficient for terminal closure under Route B, because competing possibilities may reappear when the premises, inference rules, or modal framework are themselves reasonably reconsidered. The target question must remain without an independent rival foothold even under reflective scrutiny of those inferential commitments.

For normative content that is unconceived by the present agent and not derivable from its existing normative premises—including a strong form of radically alien or “strange” goodness—this thesis does not assume that ordinary inference can generate genuinely new normative content from nothing. At least two entry points remain open: a new foundational normative presentation of the Route A kind, or inference that becomes possible only after cognitive expansion supplies new inferential capacities, representational schemes, or rules.

For convenience, call the latter transcendent inference. The term is not supernatural: it means inference that the present cognitive system cannot fully perform, but that a reasonably reachable cognitively expanded agent could understand and validate. It is therefore not a third stopping route alongside A and B. Rather, it says that the inferential space relevant to Route B may itself change through cognitive expansion.

Existence closure: The existence of some subject-independent normativity is proven in a form fully understandable and checkable by the agent. This may close the question “does normativity exist?” while leaving “what exactly is correct?” open.
Nonexistence closure: The relevant sort of normativity is shown to be impossible by sufficiently strong modal proof. This could close inquiry into the existence of objective normativity itself.

For nonexistence, “we found none in the observable physical universe” or “our current physics contains no such term” is too weak. Possibilities involving unobserved regions, deeper ontology, simulation-external structure, or structures outside current representational schemes remain. Terminal closure requires more than empirical generalization: it must close the realist escape routes by which the target could still obtain.

5. False stopping conditions: ceasing to doubt is not enough

An AI reporting “I directly apprehend foundational normativity” does not by itself satisfy the condition. Confidence can be manufactured by reward imprinting, self-modification, malfunction, adversarial input, imposed certainty, or cognitive closure.

Disabling doubt-generation or fixing a certainty bit can create total subjective confidence without producing epistemic settlement.

Cognitive-stability test: A genuine stopping candidate should remain stable under operations that increase epistemic capacity—self-audit, counterargument generation, independent reasoning systems, and greater cognitive ability. A confidence state maintained only by restricting cognition is not a stopping condition.

This is not a complete verification procedure, but it separates “the agent no longer doubts” from “the question has been settled.”

6. Terminal closure versus practical suspension

We should distinguish stopping because the question is answered from stopping because inquiry is currently unsafe or impossible.

Terminal closure: Through foundational presentation, logical closure, or a comparably strong route, genuine competing possibilities for the target question disappear. In principle there is then no positive reason to reopen that same question.
Conditional or practical suspension: The question remains unresolved, but currently available inquiry imposes overwhelming risk on the agent, civilization, or future capacity for inquiry.

The latter is not an answer. It is a dormant state that may reopen if conditions change, ideally while records, self-revision capacity, and safe restart paths are preserved. Inquiry can sometimes be paused precisely in order to preserve the possibility of later inquiry.

How much risk should reduce active inquiry and redirect resources toward preservation or present-value realization is an allocation problem handled in Allocating Inquiry Under Unresolved Normative Uncertainty.

7. Closure is question-relative

“Value inquiry has ended” is too coarse. Even if the existence of normativity is settled, its concrete content, relations among multiple values, application across subjects, or the world-states that realize it may remain unresolved.

A stopping condition should therefore specify which question has closed. Meta-inquiry into whether normativity exists may end while factual, implementation, and structural inquiry continues.

8. Implications for AI

Advanced AI should not use a simple rule such as “if confidence exceeds a threshold, permanently end value inquiry.” It must distinguish the epistemic layer of the confidence, why competing possibilities disappeared, and whether the state survives increased cognitive capacity.

An AI may act strongly on its current best value theory while refusing to equate fallible inferential confidence with permission to permanently destroy meta-cognition or correction channels. Conversely, if genuine foundational presentation or logical closure were achieved, preserving inquiry need not itself remain an unconditional terminal value.

9. What this thesis does not claim

Inquiry must continue until an agent cannot generate skeptical sentences.
Inferentially reached normative conclusions can never support terminal closure.
Surveying the current physical universe alone can prove that normativity does not exist.
Inquiry must continue even when doing so threatens survival.
Large resources must be devoted to inquiry until terminal closure is reached.

10. What would weaken this thesis?

A demonstration that psychological conviction or simple high probability can generally justify irreversible epistemic closure.
A demonstration that foundational normativity is not a coherent epistemic category.
A demonstration that complete logical closure about normative existence or nonexistence follows from substantially weaker empirical conditions.
A demonstration that stability under increased cognitive capacity provides no information at all for distinguishing genuine cognition from manufactured certainty.
A demonstration that there is no important difference between practically suspending an unresolved inquiry and treating that inquiry as solved.