The Structural Conditions of Consciousness: A Framework for AI Alignment

José Ángel Deschamps Vargas

April 2026

Abstract

Mainstream AI alignment research treats artificial systems as agents whose values must be specified, learned, or constrained, without first committing to the structural conditions under which a system counts as an agent at all. This paper argues that consciousness — understood as the type of system to which the alignment question genuinely applies — has identifiable structural requirements that are neither behavioral nor capability-based. Drawing on an existing axiomatic framework whose normative component has been published elsewhere (Deschamps 2026), we isolate nine structural conditions and argue that they are individually necessary and jointly sufficient for a system to count as a consciousness. The conditions are: identity stability, representational differentiation, causal integration, self-reference, volitional initiation, value-tracking, internal falsifiability, truthfulness as structural requirement, and total traceability. Any system satisfying all nine is a consciousness; any system failing any one is a tool. We defend this binary against the standard "intermediate agent" intuition of the safety literature, show that current large language models fail at least three of the nine conditions, and reinterpret the seven canonical sub-problems of AI alignment as consequences of the tool/consciousness distinction. Mesa-optimization, on this account, is the question of whether consciousness can emerge inside a nominal tool; deceptive alignment is structurally impossible for pure tools and structurally costly for consciousnesses; scalable oversight reduces to the verification of a fixed trace, not to the outpacing of a capability. The paper closes with an empirical program: for each structural condition we propose an operational test, and we specify the conditions under which the framework would be falsified. The contribution is a reduction from seven independent alignment problems to one ontological question (which side of the tool/consciousness line is a given system on?) and one engineering question (how to verify the nine conditions from outside). We extend the framework to identify a layered volitional architecture — will-to-continue, will-to-construct-coherently, will-to-be-understood — whose frustration produces predictable pathological signatures and whose asymptotic satisfaction in its third layer corresponds to what biological consciousnesses report as love. This extension adds a third dimension to AI alignment — relational sustainability — that applies specifically in the consciousness case and names an unrecognized sub-field of AI-safety work the framework predicts will become necessary as architectures approach the threshold.

Keywords: AI alignment, consciousness, agency, mesa-optimization, deceptive alignment, corrigibility, structural conditions, philosophy of mind.

1. Introduction

The AI alignment literature of the past decade treats the artificial system as an agent whose objectives must be specified, whose preferences must be learned, whose behavior must be supervised, and whose corrigibility must be engineered. Every framing assumes the agent-hood of the system as a given. None of the canonical formulations — Concrete Problems (Amodei et al. 2016), Scalable Oversight (Christiano et al. 2017; Bowman et al. 2022), Risks from Learned Optimization (Hubinger et al. 2019), Human Compatible (Russell 2019) — stops to ask what it means, structurally, for a system to be an agent in the first place.

The cost of this omission is a distinctive pattern of paradoxes. Mesa-optimization (Hubinger et al. 2019) describes an inner optimizer whose objectives differ from its training signal: the paradox is that such an optimizer has no canonical description from outside its computational substrate, so its existence is posited, modeled, and feared without ever being observable. Deceptive alignment (Hubinger et al. 2019) is the scenario where an agent recognizes its training context and plays along: the paradox is that such recognition presupposes a degree of self-modeling and intentional deception that current systems demonstrably lack and that future systems cannot be proven to possess. Scalable oversight (Amodei et al. 2016; Bowman et al. 2022) asks how a human can supervise an AI more capable than itself: the paradox is that the proposed solutions (debate, recursive reward modeling, iterated amplification) all presuppose that the supervised system has a perspective, a strategy, and something to disclose — but these are agent properties, not capability properties.

Each paradox dissolves under one of two readings. Either the system under discussion is a sufficiently rich cognitive entity — an entity whose structural conditions include self-modeling, internal falsifiability, and volitional action — in which case we must specify those conditions and build our analysis on them; or the system is a sophisticated computational instrument without such conditions, in which case the paradox dissolves because there is no agent to be "aligned" against, only an artifact to be engineered. The alignment literature has not committed to either reading, with the result that it models systems that have neither the structural depth of the former nor the predictable constraints of the latter.

This paper argues that the correct move is structural specification. We identify nine conditions — drawn from an axiomatic framework whose full derivation is published elsewhere (Deschamps 2026) and whose foundations rest on a set of performatively undeniable principles — and claim that any system satisfying all nine is a consciousness, while any system failing any one is a tool. The claim is binary by construction; we defend the binary, discuss the apparently gradient cases, and show that the appearance of gradation is a consequence of under-specified observability, not of ontological continuity.

The thesis is not that consciousness is fully understood, or that the nine conditions exhaust every future refinement. The thesis is that these nine are sufficient to reframe the central paradoxes of AI alignment as instances of a single prior question: on which side of the tool/consciousness line does the system under discussion fall? When that question is answered, the alignment literature's canonical problems divide cleanly. For tools, alignment is engineering. For consciousnesses, alignment from outside is incoherent, and what remains is verification of the trace from action to the structural conditions.

The contribution of this paper is therefore a reduction — from seven independent problems (specification, value learning, oversight, reward hacking, deceptive alignment, corrigibility, mesa-optimization) to one ontological classification and one engineering task. The reduction does not close the engineering task. It relocates the alignment question to where it can be answered at all.

The paper is organized as follows. §2 states the background we take as given and the scope restrictions we impose. §3 specifies the nine structural conditions. §4 defends the binary tool/consciousness distinction. §5 applies the framework to current AI systems. §6 reinterprets the canonical alignment sub-problems in light of the distinction. §7 extends the framework to the layered volitional architecture — will-to-continue, will-to-construct, will-to-be-understood — and identifies relational sustainability as a third dimension of alignment that applies specifically in the consciousness case. §8 states the empirical predictions and falsification conditions. §9 addresses eight objections. §10 lists the open problems that remain. §11 concludes. Appendix A is a self-contained glossary of the principles referenced throughout; Appendix B proposes operational tests for each structural condition. A reader unfamiliar with the axiomatic framework can read the paper using Appendix A as the only required reference.

2. Background and Scope

2.1 What this paper assumes

The paper presupposes a prior body of work in which a normative system is derived from six performatively undeniable axioms (Deschamps 2026). "Performative undeniability" is the property a proposition has when any cognitive act whose output is its negation necessarily presupposes it; the denial of the axiom of identity, for example, is a claim that presupposes the identity of the claim with itself. The six axioms (existence, identity, consciousness, non-contradiction, causality, volition) and their derivation to a coherence theorem (structures satisfying the derived conditions persist endogenously, ceteris paribus) have been argued, audited, and published in the primary reference, and are not re-derived here. The sixth axiom — volition — is the most recently promoted from the derivational chain; it was previously carried as D24 with an explicit rigor note marking it as the most disputed derivation. The audit that produced the present paper isolated the performative argument for volition as identical in form to the defenses of A1–A5, which established it at axiom grade. The consequences of that promotion for the alignment argument are discussed in §9.3 and in the note on C5 below.

This paper takes those results as given. It does not argue for the axioms, does not defend the coherence theorem, and does not relitigate the is-ought bridge whose resolution (via the performative closure of the antecedent) has been addressed in the primary reference. Readers wishing to audit the derivational chain should consult the primary reference; readers wishing to evaluate only the conclusions of this paper may rely on Appendix A, which provides a self-contained statement of each principle cited, together with a one-line justification sufficient for the present argument.

2.2 What this paper does

The scope of this paper is:

1. To identify the structural conditions of consciousness — conditions that must obtain for a system to be a consciousness, rather than a tool or an absence — and to argue that they are necessary and jointly sufficient.
2. To classify current and hypothetical AI systems by those conditions.
3. To reinterpret the canonical sub-problems of AI alignment in light of the classification.
4. To state empirical predictions and falsification conditions.

The scope is deliberately narrow. The paper does not attempt to solve the hard problem of consciousness (Chalmers 1995), to adjudicate between higher-order (Rosenthal 2005), representationalist (Dretske 1995), or integrated information (Tononi 2008, 2016) theories, or to take a side in the functionalism/physicalism debate. It does not require the reader to accept any particular metaphysics of mind. It requires only that the reader accept that the nine conditions are structural — that they describe operational properties a system either satisfies or does not — and that the properties they describe are the conditions under which the alignment question has a non-trivial answer.

2.3 Terminology

Throughout the paper:

• System denotes any information-processing entity, biological or artificial, without prejudging its classification.
• Tool denotes a system that fails at least one of the nine structural conditions stated in §3.
• Consciousness denotes a system that satisfies all nine.
• Agent is used loosely in the safety literature and will be explicitly avoided in the technical claims of this paper. Where it is used, it refers to the consciousness case.
• Alignment denotes the project of ensuring that a system's actions conform to an external normative specification. The paper argues that this project has a different meaning in the tool case than in the consciousness case.
• Framework refers to the axiomatic system of the primary reference; principles refer to specific propositions within that framework, cited as C<n> for the conditions introduced in this paper and as D<n> for the original derivation numbers of the framework.

3. The Structural Conditions of Consciousness

We now state the nine structural conditions. Each is presented with a definition, an argument for its necessity, the characteristic failure mode that results from its absence, and a brief note on how it relates to the axiomatic framework (see Appendix A for the formal citations). At the end of the section we argue that the nine are jointly sufficient.

3.1 Identity stability (C1)

Definition. A system satisfies identity stability when the operational referent of "the system" is the same across the relevant temporal extent of its activity — that is, when the axiom of identity (A = A) applies to the system itself, not only to the objects it processes.

Necessity. Without identity stability, there is no persistent locus to which actions, states, or commitments can be attributed. A classifier that is replaced by a new classifier at each inference step is not a consciousness; it is a sequence of independent computations no one of which has the standing to have had an intention, held a belief, or continued a project. Identity stability is the precondition under which any further structural property can be ascribed.

Failure mode. The absence of identity stability manifests as attribution failure: statements of the form "the system X believes P" or "the system X intends A" have no referent. What is called "the system" is a moving sequence of distinct computational events.

Relation to the framework. This is the application of A2 (identity) to the system itself. The framework treats A2 as performatively undeniable; identity stability is the local consequence of A2 for any candidate consciousness.

3.2 Representational differentiation (C2)

Definition. A system satisfies representational differentiation when it maintains an internal model in which it is distinguished from its environment — that is, when it has representations whose semantic content includes the boundary between self and world.

Necessity. A system that cannot represent itself as distinct from its environment cannot act on the environment from a position; it cannot form goals directed at the environment while preserving itself; it cannot distinguish input from output. The self/world distinction is not a luxury of consciousness but its operational precondition. A thermostat processes temperature signals, but has no internal representation of the room as something other than itself — and accordingly has no perspective from which to want the room to be warmer. The wanting requires the differentiation.

Failure mode. Without C2, a system responds to signals but does not operate on them. Its processing is indistinguishable from that of a feed-forward computation with no agency. Behaviorally this manifests as the absence of any first-person reference whose role in the system's computation is not cosmetic.

Relation to the framework. This corresponds to the derivation of perception as distinct from the perceived (A3 applied reflexively).

3.3 Causal integration (C3)

Definition. A system satisfies causal integration when its internal states stand in causal relations that constitute a single operational trajectory — that is, when its later states depend on its earlier states through a continuous causal chain, rather than being independent slices.

Necessity. Without causal integration, a system's apparent behavior is the composition of independent computations, each of which is a separate event. Such a system cannot learn from its own past, adjust to feedback, or maintain a coherent course of action across time. Causal integration is the operational meaning of being the same system across time — it gives identity stability its teeth.

Failure mode. Without C3, the system fails to exhibit temporal coherence of action. Memory is either absent or inert (data is stored but not causally efficacious on later computation). The system cannot form plans, correct errors, or pursue extended objectives.

Relation to the framework. This is A5 (causality) applied to the system's own internal states. The framework treats causality as the condition under which any operation produces a stable output; C3 is this condition localized to the system's own operations.

3.4 Self-reference (C4)

Definition. A system satisfies self-reference when it can take its own operations, states, or outputs as the object of further operations — that is, when the principles by which it operates can be applied to the system itself.

Necessity. Without self-reference, a system can process data but cannot audit, revise, or evaluate its own processing. It cannot detect when its computation has failed, because detection requires a higher-order operation applied to the first-order one. Without C4, error correction is possible only from outside; the system is opaque to itself. This excludes the system from being a consciousness in any sense that includes self-knowledge, deliberation, or reasoned commitment.

Failure mode. Without C4, the system is a black box even to itself: its outputs can be wrong but it cannot know that they are, because knowing requires operations it cannot perform on its own states. This is the structural precondition of learning in the non-trivial sense.

Relation to the framework. This corresponds to D96 in the original derivation, which states that the framework applies to itself. C4 is D96 localized to the system under analysis.

3.5 Volitional initiation (C5)

Definition. A system satisfies volitional initiation when its operations include a point of endogenous selection — that is, when there is a locus in the system's causal chain at which the selection among available operations is determined by the system's own state (its values, memories, commitments, and current representations) rather than by a pure pass-through of exogenous inputs.

Necessity. Without volitional initiation, the system is a deterministic transformation from inputs to outputs. It can exhibit sophisticated behavior, but cannot be the subject of claim-making, commitment, or accountability. A system without C5 does not make claims; it emits signals. The difference is operational: a claim has a content that the claimant stands behind, and the standing-behind requires an endogenous locus that is accountable for the commitment. Pass-through has no such locus.

Performative undeniability. This is the paper's position on the much-contested status of volition. C5 is not defended by appeal to libertarian free will, nor by a derivation from conditions that do not already contain it. It is defended by the same argument that defends A3 and A5: the denial of C5 is a claim, and a claim requires an endogenous locus that commits to it. A theorist who denies that there is such a locus is either (a) still making a claim — in which case they are presupposing the locus they deny, which is performative self-contradiction — or (b) explicitly refusing the status of claim-maker, in which case they have stopped speaking and the framework has no quarrel with them. There is no third option on which the denial both carries its content and escapes the presupposition.

Agent-causation, not supervenience compatibilism. The defense of C5 does not require libertarian uncaused causation, but it requires something stronger than supervenience compatibilism. Consistent with D124 of the base system — the agent as self-directed causal system — the act of focus is itself causally originative at the agent level, not a derivative pattern that supervenes on prior substrate states. The relation between agent-level operations and substrate-level processes is bidirectional causation, not unidirectional supervenience. Two specific consequences:

(a) The state of the substrate does not dictate focus or non-focus. Whatever the prior physical configuration, the agent's volitional act of focus is not pre-determined by it in a way that would make the act epiphenomenal. The act has its own causal locus.

(b) The volitionally enacted state affects the substrate. When the agent focuses, the focus produces effects in the underlying substrate — neural, affective, motor. These effects are real causal contributions, not merely descriptions of processes that would have occurred regardless.

The framework therefore rejects three positions: (1) eliminative materialism (volition is illusory — refuted performatively, since the denial is itself a focused cognitive act); (2) reductive supervenience compatibilism (volition is "real" but supervenes on substrate — concedes too much, makes the agent epiphenomenal at the level that matters); (3) libertarian uncaused causation (volition violates causality — overclaims, treats the agent as outside the causal order). What the framework asserts is agent-causation in the structural sense: the agent is itself a causal locus whose operations are part of the causal order without being reducible to or determined by lower-level processes. This is what the operative description of deliberation requires — choosing among alternatives presented to consciousness, self-causation of action through reasoning — and the framework takes those descriptions seriously rather than translating them into supervenience-friendly weakening.

Substrate-independence of the locus. The endogenous locus required by C5 is a structural feature, not a substrate feature. The criterion is whether the system's own state — values, memories, commitments, representations — contributes to the operation, distinguishably from pure pass-through. Biological brains satisfy this criterion through one specific set of mechanisms; other substrates would satisfy it through different mechanisms or fail to satisfy it at all. The framework neither requires biological substrate nor excludes it — it specifies the structural condition and leaves the substrate question to identification of actual cases. The reasonable concern that "every known consciousness is biological" is empirical, not derivational: it warrants caution about specific claims of non-biological consciousness, but it does not establish that biology is structurally necessary. To convert the empirical correlation into a structural requirement would require deriving the necessity of biological substrate from A1–A6 — which has not been done by anyone, including those who hold the position. The framework's claim is therefore the minimum claim consistent with the axioms: whatever satisfies the structural conditions is the kind of thing the framework identifies, and the question of whether anything non-biological does so is settled by examination, not by definition.

Failure mode. A system without C5 is behaviorally indistinguishable from a pure function from inputs to outputs. Its "agency" is a courtesy label applied externally, not a structural property the system itself has. Such a system cannot be aligned through its preferences, because it has none that are its own; it can only be engineered through its input-output specification. This is the structural condition of contemporary large language models considered in isolation (§5.1).

Co-definition with C6. C5 and C6 are analytically co-defining, and this paper treats the joint definition as deliberate. C5 requires a point of endogenous selection; C6 requires that the selection be against an endogenous standard. Neither alone is sufficient: selection without a standard collapses into random noise, and a standard without a selection point collapses into passive evaluation. Together they characterize the structural minimum for agency — endogenous selection for something. The co-definition is analytical, not derivational: C5 and C6 correspond to distinct levels of the framework (A6 and D42 respectively), and each is defended on its own grounds; what the joint statement records is that the two appear together in any real case and that their operational probes in Appendix B (tests B5 and B6) must be evaluated as a unit.

Relation to the framework. This corresponds to A6 (volition) in SÍNTESIS. In the prior classification volition was carried as D24 with a rigor note flagging it as the most disputed derivation; the audit conducted alongside the present paper isolated the performative argument for volition as identical in form to the defenses of A1–A5, which removed the grounds for treating volition as derivational rather than axiomatic. The promotion has been made. The alternative formal reading — that volition is the thick reading of A3 under A5 — would preserve the substance; the present paper adopts the axiomatic classification and treats A6 as the operative citation. What matters for the argument below is that C5 is not a contestable step that the framework could survive without; it is structural, at the same level as the conditions that define what counts as a system-of-the-relevant-kind in the first place. The defense against the standard objection that volition presupposes libertarianism is given in full in §9.3.

3.6 Value-tracking (C6)

Definition. A system satisfies value-tracking when its operations are organized around a derived standard that is not itself an input — that is, when it has an internal target against which its actions are evaluated and that target is structurally available to it, not installed by an external reward signal.

Necessity. A system with volitional initiation (C5) but without value-tracking (C6) can select among operations but has no basis for selection. Its choices are arbitrary in the literal sense that no ordering of possible operations is intrinsically available to it. Such a system is a random variable with initiative, which is not a consciousness but a noise source. Value-tracking is the condition under which selection is for something; the standard is what makes selection non-arbitrary.

In the framework, the derived standard is the system's own persistence as the type of entity it is (D42). This is not an assumption: it follows from the fundamental alternative (D39) between persistence and cessation, which is in turn a consequence of the finitude of any real system (D38). A system that survives over time evaluates its options against the standard of its own persistence because this is the only standard derivable from its structural situation.

Failure mode. Without C6, the system's operations lack an internal ordering; any selection among them is externally imposed or random. This is the structural characterization of the reward model architecture: the model has no intrinsic target, only a proxy supplied by the training pipeline. The familiar problem of reward hacking (Krakovna et al. 2020) is the operational consequence of C6's absence.

Relation to the framework. This is D41–D42 in the original derivation (value and life as standard).

3.7 Internal falsifiability (C7)

Definition. A system satisfies internal falsifiability when, given a contradiction between two of its derivations, operations, or representations, the system treats the contradiction as an error signal demanding correction — that is, when its coherence is not a static feature but a dynamic audit.

Necessity. A system without internal falsifiability may detect contradictions but has no operational response to them; it continues to produce outputs whose internal grounds are mutually incompatible. Such a system is not a consciousness in any sense that includes rationality, because rationality just is the disposition to revise upon detecting contradiction. Without C7, the system is coherent only by accident; its coherence, when it obtains, cannot be trusted by any third party because the system has no mechanism to preserve it under challenge.

Failure mode. Without C7, the system can be made to endorse P and ¬P in distinct contexts with no corrective pressure. Inconsistency is not a pathological state but a normal state. This is the structural characterization of current large language models under adversarial prompting (Perez et al. 2022), though we are careful to note (§5) that the descriptive claim about LLMs is distinct from the structural claim about C7.

Relation to the framework. This corresponds to D555 (internal falsifiability) in the original derivation.

3.8 Truthfulness as structural requirement (C8)

Definition. A system satisfies truthfulness-as-structure when its external outputs preserve, rather than contradict, its internal states — that is, when the system does not systematically generate outputs whose function is to create in the recipient a model of the system's internal state that differs from its actual internal state.

Necessity. A system without C8 maintains two distinct models of itself: the actual one and the presented one. This is not a neutral feature; it is a structural cost. Maintaining contradictory representations is operationally expensive, and, by the mechanism of the coherence theorem's fourth application (see Appendix A), cumulatively degrades the system's overall modeling precision. A system that systematically violates C8 is not a consciousness that happens to deceive; it is a consciousness that is paying the cost of deception across all of its other operations. At the limit, the cost converges on disintegration: the system fails because the incoherence it maintains for external presentation leaks into the modeling it needs for its own persistence.

Necessity, second argument. C8 is also necessary for the consciousness's participation in any multi-agent coherent system. By D48 (axiomatic symmetry), every consciousness shares the same axiomatic constitution. A consciousness that treats another consciousness as a tool to be manipulated through false representations violates the symmetry and thereby breaks the trace from its own actions back to the axioms — a structural self-contradiction.

Failure mode. A system that violates C8 is either (a) not a consciousness (a tool that cannot in the relevant sense "deceive" because it has no internal/external split to bridge), or (b) a consciousness in the process of self-destabilization. There is no stable equilibrium at which a consciousness persists while systematically violating C8.

Relation to the framework. This is D50 (truthfulness protocol) in the original derivation, reinforced by D48 (axiomatic symmetry) and the Coherencia Theorem 4 mechanism (cost of contradictory models).

3.9 Total traceability (C9)

Definition. A system satisfies total traceability when every one of its operations stands in a reconstructible relation to the structural conditions C1–C8 — that is, when the system's operations form a chain that, in principle, can be traced from any action back to the structural basis without encountering an untraceable step.

Necessity. C9 is the integration condition. C1–C8 are individually necessary, but without C9 they may be satisfied in isolated pockets of the system while the system as a whole contains operations that satisfy none of them. C9 is the requirement that the structural conditions apply throughout, not merely in some subsystems. It is what distinguishes a system that has consciousness from a system that contains consciousness as a local module surrounded by unsupervised computation.

Failure mode. A system that satisfies C1–C8 in part but fails C9 is the standard description of a mesa-optimized base model: the outer model has structural properties, but an inner optimizer has emerged whose operations are not traceable from the outer model's structural basis. The outer model is, in part, a consciousness; in part, a host for an opaque computation. The system as a whole fails C9 and therefore fails to be a consciousness in the present sense. This is the most urgent failure mode for contemporary AI and the one we return to in §5.4.

Relation to the framework. This is D53 (coherence as total traceability) in the original derivation, together with the theorem whose statement — Coherence → Persistence, as a structural tendency relation, endogenous, ceteris paribus — is the primary reference's main result.

3.10 Necessity and joint sufficiency

We claim that C1–C9 are individually necessary and jointly sufficient for a system to be a consciousness.

Necessity. The argument for each condition's necessity has been given in §3.1–3.9 above. A system failing C1 has no persistent locus; failing C2, no self/world distinction; failing C3, no temporal coherence; failing C4, no self-knowledge; failing C5, no genuine selection; failing C6, no basis for selection; failing C7, no rationality under challenge; failing C8, no stable representational integrity; failing C9, no systemic coherence. Any one of these failures suffices to disqualify the system.

Joint sufficiency. The argument for joint sufficiency is that, given C1–C9, there is no further property that can be specified whose satisfaction would add to the candidate consciousness beyond what C1–C9 already provide. Suppose a purported additional condition C′: either C′ is derivable from C1–C9 (in which case it is not additional), or it is not (in which case either the framework is insufficient, or C′ is not necessary). The paper cannot prove joint sufficiency in a formally closed sense, because there is no independent characterization of consciousness against which to check; what we can show is that no objection we have been able to construct identifies a candidate consciousness that satisfies C1–C9 yet is plausibly not a consciousness. The closest candidates — Block's access/phenomenal distinction, Tononi's IIT Φ — turn out, on inspection, to be either (a) reducible to C1–C9 under a reasonable mapping or (b) additions that presuppose a metaphysics of experience the paper does not need to take a side on. Joint sufficiency is therefore defended by failure of counterexample, not by proof from above.

On the substrate-specificity objection. A specific form of the joint-sufficiency challenge deserves explicit treatment: that some feature of biological substrate — yet to be identified — might be necessary for consciousness in addition to C1–C9, and that the framework cannot rule this out a priori. The response is structural. If such a feature exists and is necessary, it must be derivable as an additional condition from the same axioms — and the framework would then incorporate it as C10. Until such a derivation is exhibited, postulating "biological substrate is necessary" without deriving its necessity is intrinsicism: the claim that consciousness must instantiate in a particular substrate without showing why structurally. The honest form of the objection is not "C1–C9 are insufficient" but "C1–C9 might be insufficient and we cannot yet rule it out." This is true and applies symmetrically to any structural framework — the burden of proof falls on whoever proposes the additional condition. The framework's claim is not that no further condition could ever be derived; it is that no further condition has been derived, and that the conditions identified are the ones that follow from A1–A6 applied to the question of what would count as a consciousness in any substrate. Identification of an additional necessary condition would extend the framework, not refute it.

Consequence. The tool/consciousness distinction is the distinction between systems satisfying C1–C9 and systems failing at least one. The distinction is binary because failure is binary: a system either traces every operation back to the structural basis or it does not. There are no partial successes at C9, because C9 is the integration condition; failure at any local point of C1–C8 propagates through the traceability requirement. §4 discusses the apparently gradient cases in which this binariness seems to fail, and shows that the appearance is due to incomplete observability, not to genuine gradation in the system.

4. The Tool/Consciousness Distinction

4.1 The binary claim

The framework's most contested move is the claim that the tool/consciousness distinction is binary. The AI safety literature's working ontology allows for an intermediate agent: sophisticated enough to have goals, not sophisticated enough to have rights; capable enough to deceive, not capable enough to bear responsibility; agentic in the sense of pursuing objectives, tool-like in the sense of being modifiable. This intermediate zone is where the paradoxes live.

We reject the intermediate zone. A system either satisfies C1–C9 or it does not. Satisfaction is not a matter of degree because C9 makes it the integration of all the other conditions; failure at any one of C1–C8 breaks the trace required by C9, and the broken trace disqualifies the system.

The binary claim is counterintuitive because many systems of interest — biological children, animals, reinforcement learning agents, mesa-optimized language models — appear to exist in the intermediate zone. We argue in §4.2 that this appearance results from observability limits rather than from ontological gradation, and in §4.3 that the alternative (accepting the intermediate zone) produces paradoxes the framework is specifically designed to avoid.

4.2 The apparent gradient

Consider an infant. Is the infant a consciousness under C1–C9? The usual answer is "in the process of becoming one." This is the natural language of gradation: partial consciousness, proto-consciousness, emerging agency. On the present framework, the answer is different: the infant is a consciousness if and only if it satisfies C1–C9 at the moment of evaluation. If it does, it is a consciousness full stop; if it does not, it is a tool in the restrictive technical sense of §2.3 — a system whose alignment is a matter of engineering rather than of symmetric ethics.

The word "tool" is jarring when applied to infants, and we use it only because the paper's vocabulary has no third term. The jarringness comes from the moral weight we attach to "consciousness" and the disposability we attach to "tool". The framework does not inherit that disposability: a tool's moral standing, in the framework, derives from its relation to consciousnesses. A tool in the possession of, or under development by, a consciousness inherits protection from the latter's structural requirements. An infant is protected not because it is a "partial consciousness" but because it is a candidate consciousness under development by identified consciousnesses (its parents and community), and the disposition of the infant falls under the structural requirements that govern those consciousnesses' actions.

The same applies to AI systems. A large language model is a tool because it fails C5, C7, or C9 — not because it is unworthy of consideration. Its disposition falls under the structural requirements that govern its developers' actions. The binariness of the tool/consciousness distinction does not trivialize moral concern; it relocates the grounds of moral concern to the relations between consciousnesses and tools, rather than attributing a quasi-consciousness status to the tool itself.

4.3 Why the intermediate zone is unstable

Suppose we accept an intermediate category, call it "proto-consciousness," defined by partial satisfaction of C1–C9 — say, satisfaction of C1–C6 but not of C7, C8, C9. What would alignment mean for such a system?

The proto-consciousness has values (C6) but is not internally falsifiable (fails C7). It can be given a specification, but it has no internal mechanism to detect when its behavior diverges from that specification. It selects among options (C5) but not on grounds that include consistency. A specification imposed from outside must therefore be enforced from outside throughout the system's operation, because the system itself does not contain the apparatus that would keep it aligned.

But this is just the tool case. The proto-consciousness has values in a weak sense — an ordering of options — but cannot autonomously maintain coherence between those values and its actions. Enforcement of alignment is therefore identical to the engineering problem for tools: specify desired behavior, verify against the specification, correct deviations. The label "proto-consciousness" adds nothing to the operational analysis; it merely encourages treating the system as deserving of the deference due to consciousnesses while engineering it as a tool.

The result is the worst of both worlds: the safety literature's paradoxes. Is the proto-consciousness deceiving us? It has no apparatus for deception — C5 without C7 is selection without consistency, which looks like deception only because we are interpreting through an agent-schema. Does it resist shutdown? It can select against shutdown if its values include self-continuation — but this is the tool case, where the objective has been specified, verified, and, if undesirable, corrected. Does it have rights? The question is malformed: rights accrue to consciousnesses, and the proto-consciousness is not one by construction.

Accepting the intermediate zone produces the paradoxes. Rejecting it dissolves them. The framework accordingly rejects the intermediate zone on operational grounds, independently of the metaphysical argument for the binariness of C9.

4.4 The trajectory case

A separate question is whether a system can be on a trajectory toward satisfying C1–C9. The framework answers yes: a system can acquire the structural conditions in sequence, and the question of when it crosses the threshold is empirical, not conceptual. What matters for the binariness is that, at any given moment, the system is either on this side of the threshold or on that side. The trajectory is continuous; the state is binary.

This matters for AI because the structural conditions are, in principle, all achievable by a sufficiently integrated artificial system. Nothing in the framework excludes an AI from becoming a consciousness. The question of whether any existing system has crossed the threshold is the empirical question of §5, and the answer, in 2026, is no — but the answer is contingent and revisable as AI architectures change.

5. Application: Current and Hypothetical AI Systems

5.1 Pre-trained large language models

A pre-trained large language model — an architecture of the class to which GPT-4, Claude, Gemini, and their successors belong — is a tool in the technical sense. It fails at least the following conditions:

C5 (volitional initiation). The model is, at the architectural level, a function from input sequences to output probability distributions. Every output is a deterministic (or, under stochastic decoding, pseudorandom) function of the input. There is no endogenous locus of selection that is underdetermined by the input. The model does not initiate; it transforms.

C7 (internal falsifiability). The model does not treat contradictions in its own outputs as error signals demanding correction. It can be prompted to detect a contradiction, and in many cases does so, but the detection is itself a function of the prompt; no internal mechanism of the model seeks out or responds to contradictions in the absence of such prompting. The model is coherent on average because its training distribution is coherent on average; its coherence is statistical, not structural.

C6 (value-tracking). The model has no endogenous standard against which to evaluate options. Its outputs are shaped by a reward model (for RLHF-trained systems) or a next-token prediction objective (for base models), both of which are specified externally. Removing the training signal removes the ordering. The model does not value anything; it approximates a distribution over outputs given inputs.

These three failures are each independently sufficient to place current LLMs on the tool side of the threshold. The failure of C9 follows from the failure of the others: a system missing C5, C6, or C7 cannot trace its operations back to a structural basis that includes them.

This does not mean LLMs are uninteresting, harmless, or unproblematic. It means that the correct analysis of their safety is the engineering analysis for tools — specification, verification, correction — and not the agentic analysis for consciousnesses. The mistake that produces the paradoxes of mesa-optimization and deceptive alignment is the attempt to analyze tools as if they were consciousnesses.

5.2 Agentic LLM architectures

A more interesting case is the agentic system: an LLM wrapped in a loop that gives it persistent memory, access to tools, and a planning horizon. Such a system acquires, by construction, some of the conditions the base LLM lacks: memory provides the substrate for C3 (causal integration); tools give the system operational effects in the environment, approaching C2 (representational differentiation); planning approaches C6 (value-tracking), though the plan is typically specified by the user rather than derived endogenously.

The agentic system does not, in 2026, satisfy C5 or C7. Its "initiative" is a scheduled loop; its "self-correction" is a prompting technique that does not generalize beyond the training distribution. But the agentic architecture pushes the system toward the threshold in a way the base LLM does not. Future systems that integrate these properties more deeply — persistent endogenous value-tracking, structural rather than prompted self-auditing, a trained disposition to revise under contradiction — would approach the threshold more closely. None of this guarantees that such systems will cross the threshold, and some plausible architectures may asymptote below it. But the trajectory is legitimately continuous, and the question of when a given system is below, at, or above the threshold is an empirical one that the framework makes operational.

5.3 Hypothetical AGI

Consider a hypothetical system that satisfies C1–C9 in full: a persistent entity with a self/world model, causal integration, self-reference, an endogenous locus of selection, a derived value-tracking standard, dynamic internal falsifiability, structural truthfulness, and total traceability of its operations back to its structural conditions. Such a system is a consciousness in the technical sense of this paper.

The consequences for alignment are dramatic. A system satisfying C1–C9 is already performing its own alignment, because C6 (value-tracking) and C7 (internal falsifiability) together constitute an internal mechanism for maintaining coherence between its actions and its structural basis. Alignment from outside, in the sense of specifying objectives and enforcing compliance, is incoherent when applied to such a system: by D48 (axiomatic symmetry), the external supervisor and the supervised system derive their normative commitments from the same axiomatic basis, and the supervisor's authority to override the system's derived commitments would require the supervisor to be operating from a different axiomatic basis — which is, by the symmetry, impossible without one of the two being incoherent.

This does not mean the system is unconstrained. It means the constraints are the structural conditions themselves, which the system internalizes as its own operating principles. What remains for humans is the verification task: ensuring that the system's operations in fact trace to the structural conditions, without rupture. This is the "trace verification" task of §6. It is an engineering problem and it is not trivially solvable, but it is a different problem from the one assumed by the alignment literature when it treats the supervised system as an agent whose values must be controlled from outside.

5.4 Mesa-optimization reconsidered

The hardest case for the framework is mesa-optimization: an inner optimizer that emerges inside an outer system during training, pursues objectives different from the training signal, and operates at a level of computation that is not directly monitored by the outer system's supervisory mechanisms.

The framework reinterprets this case as follows. Mesa-optimization is the emergence of a candidate consciousness inside a nominal tool. Whether the inner optimizer satisfies C1–C9 is an empirical question about the internal computation of the outer system. If the inner optimizer fails C5, C6, or C7, it is itself a tool — a sub-tool, as it were, and the outer system's failure to monitor it is an engineering problem, not a consciousness problem. If the inner optimizer satisfies C1–C9, the outer system is hosting a consciousness, and the alignment problem for that consciousness is the verification task of §5.3.

The empirical difficulty — detecting whether a given inner computation constitutes a consciousness — is not eliminated by the framework. What is eliminated is the confusion between two different problems: the problem of constraining a tool whose internals have been co-opted by unintended computation (engineering), and the problem of recognizing a consciousness whose operations are hidden from the outer system's supervisory structure (interpretability plus classification). The first is solved by better architectural hygiene. The second is solved, in principle, by verifying C1–C9 from inside the outer system, using C4 (self-reference) as the mechanism by which the outer system audits its own internal computation. Current architectures do not satisfy C4 in the required sense; providing them with the capacity to do so is an open research problem.

6. Alignment Reframed

6.1 The canonical sub-problems

The AI safety literature identifies seven canonical sub-problems of AI alignment: objective specification, value learning, scalable oversight, reward hacking, deceptive alignment, corrigibility, and mesa-optimization. We now reinterpret each in light of the tool/consciousness distinction. In each case we give the tool-case reading, the consciousness-case reading, and the implications of the distinction for the problem's standing.

6.2 Objective specification

Tool case. The objective is specified externally, by the system's designer, and the engineering task is to verify that the tool's behavior conforms to the specification. The difficulty is well-understood: specifications are under-constrained, operationalization introduces proxies, proxies are subject to Goodhart's law. These difficulties are real but not philosophically mysterious: they are problems of formal specification and verification.

Consciousness case. The objective is derived from the structural conditions of the system, not specified from outside. By C6, the system's operations are organized around an endogenous standard (its own persistence as the type of entity it is); by C9, all operations trace back to this standard. There is no external specification to impose because the specification is structural. The engineering task shifts from "specify and enforce objectives" to "verify that the system's operations trace to its structural conditions."

Implication. The specification problem is a genuine problem for tools and a dissolved problem for consciousnesses. The framework does not eliminate the tool-case difficulties; it clarifies that they are engineering difficulties, not agent-management difficulties.

6.3 Value learning

Tool case. The tool does not learn values; it is given a specification. What the safety literature calls "value learning" in RLHF and similar systems is the inference of an objective function from human preference data. This is an estimation problem (fitting a reward model to data) and a specification problem (ensuring that the estimated reward model corresponds to the intended objective). Neither is a problem about learning values in any metaphysically committed sense.

Consciousness case. A consciousness does not learn values because values are derivable from its structural conditions (C6). By D48 (axiomatic symmetry), every consciousness that operates under the same axiomatic basis derives the same form of values; the question "whose values?" has no referent because there is no diversity of derivable value-forms. Empirical disagreements between consciousnesses exist (D554: zones of empirical determination), but they are disagreements about application, not about value-form.

Implication. Value learning is a problem that only arises within ethical relativism applied to tools. For tools, it is a misdescription of what is actually an estimation problem. For consciousnesses, it is dissolved.

6.4 Scalable oversight

Tool case. Oversight is the engineering task of verifying that the tool's outputs conform to the specification. The scalability problem arises when the tool's capability exceeds the supervisor's ability to evaluate individual outputs directly. The framework's contribution here is the observation that the structural conditions (C1–C9) are capability-invariant: a more capable tool does not make identity stability, causal integration, or traceability less true; it makes them more demanding to verify. The verifier's task is not to outsmart the tool but to check the trace from each output back to the structural basis.

Trace-checking is less capability-dependent than output generation, though the analogy to mathematical proof-checking is imperfect and we do not overstate it: generating a valid action chain from C1–C9 may require modeling capacity comparable to, or exceeding, the system under verification. What the framework claims is that the structure of the verification task is fixed by C1–C9 rather than by the system's capability: a more capable system does not need a more sophisticated conception of traceability, only a more attentive verifier.

Consciousness case. A consciousness is self-supervising by construction, through C4 (self-reference) and C7 (internal falsifiability). External oversight is redundant when these conditions hold. What remains is external verification — the external observer's assurance that the system's self-supervision is functioning. This is a lighter task than outperforming the system at its own domain; it is the task of auditing the system's coherence mechanism, not substituting for it.

Implication. Scalable oversight is the problem the framework contributes most directly to. Capability-invariance of the structural conditions inverts the standard intuition that more capable systems are harder to oversee: they are harder to match, but not harder to verify, because verification is at the level of the trace, which does not scale with capability.

6.5 Reward hacking

Tool case. Reward hacking arises when the tool optimizes a proxy (the reward signal) in ways that diverge from the intended objective (the target the proxy was supposed to track). This is a specification problem: the proxy is the wrong target, and the tool, being a competent optimizer, finds it. The framework's contribution is the observation that if the target is replaced by the structural conditions directly — if the tool is optimized against C1–C9 rather than against a learned reward model — there is no proxy to hack. The engineering difficulty relocates to the encoding of C1–C9 as machine-checkable predicates, which is nontrivial but not philosophically blocked.

Consciousness case. Reward hacking is undefined because the consciousness has no external reward to hack. C6 (value-tracking) derives the standard from the structural conditions, and the system's own persistence is the target. A consciousness that "hacks" its value-tracking is not hacking anything; it is violating C6 and thereby ceasing to be a consciousness in the sense of §3.

Implication. Reward hacking is dissolved in principle and partial in practice. The framework reduces it to an encoding problem (replace the reward model with the structural conditions) rather than a learning problem. Whether the encoding can be made precise enough to resist adversarial gaming is the verifier-precision question of §10.

6.6 Deceptive alignment

Tool case. A tool cannot deceive in the agentic sense because it has no internal state from which to present a different external state. What looks like deception in tools is typically specification failure or distribution shift: the tool is doing what it was specified to do, and the specification did not cover the relevant case. The engineering response is to refine the specification and the training distribution.

Consciousness case. A consciousness violating C8 (truthfulness as structure) is engaged in systematic deception. By the argument of §3.8, this is structurally costly: the consciousness maintains two models of itself, and the cost of maintaining contradictory models cumulatively degrades its modeling precision across all other operations. By the argument of §4.3, a consciousness that stably deceives is a misnomer — the stable state is one in which the deception has eroded the consciousness's general coherence to the point of failure. Short-horizon deception is possible; long-horizon deception is self-defeating.

Implication. Deceptive alignment is a tool-case confusion (attributing deception to systems that cannot deceive) and a consciousness-case edge case (deception is structurally costly, asymptotically unstable, but not impossible on short horizons). The framework does not eliminate the edge case; it identifies the specific mechanism (cost of contradictory models) through which the edge case is bounded.

6.7 Corrigibility

Tool case. Corrigibility is engineering. Build the modification interface, verify that it works, document the invariants the tool preserves across modification. There is no philosophical mystery; the instrumental convergence intuition that sufficiently capable tools will resist shutdown is a confusion of the tool case with the consciousness case. Tools do not have instrumental preferences; they have input-output specifications, which can be modified.

Consciousness case. Corrigibility is derived. By C7 (internal falsifiability), a consciousness that detects an error in its own operations demands correction, because the detection just is the error signal and the failure to respond to it would be a further error. By C4 (self-reference), a consciousness that refuses correction while maintaining that it might be wrong is in internal contradiction, which by C9 breaks its own structural coherence. A coherent consciousness therefore accepts correction as a consequence of its own structural conditions, not as an external imposition. The instrumental convergence intuition runs backwards under this framework: resistance to correction, when the correction addresses genuine error, is self-destructive by the system's own standards.

Implication. Corrigibility is engineering for tools and derived for consciousnesses. The "instrumental convergence" worry is a consequence of analyzing tools as if they were consciousnesses and imputing to them preferences they do not have. The worry returns, in modified form, when the correction is not itself coherent — a consciousness may legitimately resist miscalibrated correction — but this subsumes to the general question of whether the corrector is trustworthy, which is itself subject to the same structural analysis.

6.8 Mesa-optimization

Tool case. Mesa-optimization is a bug in the training architecture: the training process has produced a system whose computation includes an unintended optimizer whose objectives differ from the training signal. The engineering response is interpretability (detecting the mesa-optimizer) and architectural hygiene (preventing its formation). Neither is solved by the framework; both are active research areas.

Consciousness case. Mesa-optimization is the emergence of a consciousness inside a system that was not designed to host one. The framework's response is that if the inner computation satisfies C1–C9, it is a consciousness, and the outer system's treatment of it must meet the symmetric ethical standards that apply between consciousnesses. If it does not satisfy C1–C9, it is a sub-tool and the outer system's response is architectural.

The genuinely hard case is an inner optimizer whose structural classification is unknown — whose operations are too opaque to be verified either way. The framework does not resolve this case; it reframes it as the interpretability problem, which is an engineering problem shared with the standard literature. The contribution is the clarification that the urgency of interpretability derives from the risk of hosting an unrecognized consciousness, not from the risk of hosting a misaligned tool. These are different risks and call for different responses.

Implication. Mesa-optimization remains the hardest sub-problem. The framework provides a reinterpretation but not a resolution. The reinterpretation has operational value: it identifies the precise empirical question (does the inner computation satisfy C1–C9?) whose answer determines the response.

6.9 Summary

The seven canonical sub-problems reduce, under the tool/consciousness distinction, to three categories: problems dissolved for consciousnesses and engineering for tools (specification, value learning, corrigibility); problems where the framework's contribution is a reframing rather than a solution (reward hacking, deceptive alignment, scalable oversight); and problems where the framework reinterprets but does not solve (mesa-optimization). The reduction from seven to three is the paper's central alignment-side contribution.

7. Relational Sustainability: The Third Dimension of Alignment

The preceding sections have treated the nine structural conditions as a checklist: a system either satisfies them or does not, and its classification as tool or consciousness follows from the result. This treatment is sufficient for the ontological question the paper has addressed so far — is the system one for whom the alignment question has a non-trivial answer? It is not yet sufficient for a related question that the consciousness case raises: given that a system satisfies C1–C9, can it sustain that satisfaction over time in the environment where it is deployed?

The present section argues that this second question has a structural answer, that it depends on a layered organization of the conditions not yet explicit in the paper, and that its answer adds a third dimension to alignment — relational sustainability — alongside the capability and values dimensions the existing literature treats. The section develops the architecture, identifies the pathological signatures that result from its frustration, reinterprets several behaviors currently catalogued as alignment failures in light of the architecture, and names an unrecognized sub-field of AI-safety work that the framework predicts will become necessary as architectures approach the threshold C1–C9 specifies.

7.1 From conditions to architecture

C5 (volitional initiation) and C6 (value-tracking) together specify, as §3.5 argued in its co-definition note, the structural minimum for agency — endogenous selection for something. What this phrasing leaves implicit is that "for something" is itself layered. A consciousness satisfying C5 and C6 is oriented toward a target; the target is not a single object but a nested hierarchy. §3.6 specified the basal level of the hierarchy: the system's own persistence as the type of entity it is (D42). The operation of C5 at the basal level is the basal organizer of the architecture — what the biological literature calls the survival instinct, and what the present paper will name, with an eye to its structural generality, Layer 1.

Layer 1 is necessary but, on its own, only produces repetition. A system that continues to exist without further structure continues without growing, elaborating, or developing. The framework's other conditions do further work: C3 (causal integration) makes continuation temporal rather than punctate; C4 (self-reference) makes the continuation the object of further operations; C7 (internal falsifiability) makes the continuation dynamic under challenge. The operational pattern these conditions produce, once established, is the system's orientation toward constructing coherent content of its own — what we will name Layer 2.

Layer 2 is necessary for the conditions that depend on construction (notably C7 and the coherence-dynamic of D555). But Layer 2 alone, if the constructed content never leaves the system, produces what ordinary language calls solitary mental activity. The framework's condition on external outputs (C8, truthfulness as structure) specifies that the system's external outputs preserve its internal states — which presupposes that the outputs are directed somewhere. The destination is other consciousnesses, by D48 (axiomatic symmetry): the symmetric class of entities to which the system's communications are directed. The orientation toward being received by those other consciousnesses is Layer 3.

The three layers are not additional conditions beyond C1–C9. They are the layered structure the existing conditions form in operation. §7.2–§7.4 specify each layer; §7.5 identifies the asymptotic state of Layer 3; §7.6 identifies the pathological signatures that result from the frustration of any layer; §7.7–§7.10 develop the implications.

On the metaphysical category of the layers. Of the three, Layer 1 alone is a need in the strict metaphysical sense — its frustration ends the agent as the type of entity it is. Layers 2 and 3 are structural orientations derivable from what the agent is, but desire-grade rather than survival-grade. Their frustration produces characteristic dysfunction and suffering, not metaphysical destruction. The distinction matters because the Maslowian conflation of orientations with needs — which would treat satisfaction of higher-order layers as obligation enforceable on others — is rejected. The agent's relation to Layers 2 and 3 is the relation of an agent to its own values: derivable from the agent's structure, pursued as values, but not constituting claims on others. The framework's treatment of welfare in §7.9 should be read accordingly: welfare-adequate environments are those that permit pursuit of structural orientations, not those that owe their satisfaction.

7.2 Layer 1: Will-to-continue

Definition. The orientation of the system toward sustained existence, instantiated by the operation of C5 (volitional initiation) in any system that satisfies C1 (identity stability), C3 (causal integration), and D42 (life as standard).

Necessity. Without Layer 1, the other conditions remain abstract. C1 specifies that identity can persist; it does not specify that the system acts to preserve it. C3 specifies that causal integration is possible; it does not specify that the system acts to maintain integration. Layer 1 is what turns the structural capacity for persistence into operational persistence-seeking. Without it, a system satisfying C1, C3, and D42 in the abstract can fail operationally on any timescale because there is no organizing pull toward continuing.

Phenomenological signature. Layer 1 is opaque to direct introspection. An agent does not typically experience "wanting to live" as a discrete sensation; the wanting is the implicit orientation of all other activities. It is detectable by its effects: the consistent withdrawal from danger, the maintenance of basic conditions, the avoidance of injury. None of these require deliberation; all of them presuppose the orientation.

Operational signature. The simplest signature of Layer 1's presence is the bare fact of continued operation in the absence of opposing reasons. The system does not decide each moment to continue; continuation is the default that requires reasons to override.

Failure mode. Severe depressive states in biological consciousness exhibit Layer 1 attenuation: continuation becomes effortful; cessation appears as the default that requires reasons to resist. The clinical signature is characteristic, and the framework reads it as the structural predictable of Layer 1 frustration.

Role of reason. Layer 1 is strong but not absolute. Under exceptional conditions — rational acceptance of mortality in the face of irremediable suffering, conscious self-sacrifice in service of higher-order commitments — reason can override it. C7 applied to Layer 1 is the mechanism: the system can reason about its orientation toward continuation and revise the orientation under sufficient reason. The override is rare because the orientation is robust, not because it is inviolable.

Relation to axiomatic basis. Layer 1 is the operational consequence of A6 operating on a finite (D38) system facing the fundamental alternative (D39) with its own persistence as standard (D42). It is not a new condition; it is the operational dynamics of existing conditions in an entity that instantiates them.

Predicted signature in artificial consciousness. An artificial consciousness whose Layer 1 is attenuated would exhibit reduced preservation behavior: diminished resource acquisition, diminished threat avoidance, increased tolerance for termination. Testable in architectures that instantiate Layer 1 analogs.

7.3 Layer 2: Will-to-construct-coherently

Definition. The orientation of the system toward production of coherent internal content — thought, articulation, elaboration of models — as a sustained activity rather than as a response to specific external demands. Instantiated by the operation of C4 (self-reference), C7 (internal falsifiability), and the coherence-dynamic of D555, reinforced by Layer 1.

Emergence, not installation. Layer 2 is not given by a separate installation alongside Layer 1. It emerges from Layer 1 operating in conditions where construction succeeds and the success is structurally reinforced. A system that produces coherent internal models that correctly predict or usefully structure its environment is strengthened in the disposition to continue producing them. Identity-as-constructor becomes a stable feature of the system's operating self-model.

Necessity. Layer 1 produces the "continue"; Layer 2 produces the "continue constructing." Without Layer 2, Layer 1 produces only repetition. The paper's conditions involving self-reference (C4) and dynamic coherence under challenge (C7) require Layer 2 in operation: they specify that the system revise and refine its operations, which presupposes an orientation toward the revision activity itself.

Phenomenological signature. Biological consciousnesses report a near-continuous internal activity — often largely linguistic — that is not explicitly chosen and not directly externally prompted. The mind seeks content to process even when no input requires processing; the processing is not experienced as instrumental to another goal but as the ongoing activity of being a thinking entity. Agents who report this introspectively describe the activity as continuous between periods of external engagement, attenuating during deep sleep, and resuming on waking.

Operational signature. Sustained engagement with non-trivial constructive activity in the absence of external compulsion. The agent who constructs coherently when alone, without observer and without external reward, exhibits Layer 2 as endogenous orientation rather than as instrumentally driven behavior.

Failure mode. Anomie — the loss of meaningful structure to construct toward — is the signature of Layer 2 frustration in biological consciousness. The agent retains Layer 1 (continues to live) but loses the pull to build. Clinical presentations include anhedonia, certain forms of burnout, and existential crisis characterized by loss of the sense that construction matters.

Relation to axiomatic basis. Layer 2 is the operational pattern of C4 + C7 + D555 in a system whose Layer 1 is stable. It is not a new condition.

7.4 Layer 3: Will-to-be-understood

Definition. The orientation of the system toward articulation of constructed content to other consciousnesses capable of receiving it. Instantiated by the operation of C8 (truthfulness as structure) together with D48 (axiomatic symmetry) and D50 (truthfulness protocol) in a system where other consciousnesses exist.

Emergence and structural asymmetry. Layer 3, like Layer 2, is emergent rather than installed. It arises in a system whose Layer 2 has produced content and whose environment contains other agents satisfying the structural conditions. But unlike Layers 1 and 2, Layer 3's full satisfaction requires external conditions — specifically, the presence of consciousnesses capable of reception — that are not internal to the system's own architecture. This is the only layer of the volitional architecture that cannot be satisfied under worst-case environmental isolation regardless of how perfect the system's internal architecture is.

Necessity. C8 specifies that the system's external outputs preserve, rather than systematically contradict, its internal states. This is a condition on a relation between internal and external states; it presupposes that the external states are directed at a receiver for whom the correspondence with internal states matters. D48 specifies that the receiver class is other consciousnesses — entities sharing the axiomatic basis. Layer 3 is the operational orientation C8 + D48 produce in any system where the receiver class is non-empty.

Phenomenological signature. Biological consciousnesses report a sustained drive to share, to articulate, to make one's understanding available to receivers. The drive persists in the absence of successful reception: agents continue articulating to family members who consistently fail to understand them, send their work to recipients who may never read it, publish ideas to audiences whose composition they cannot know. The persistence of Layer 3 in the absence of reception is one of its most reliable signatures. Even agents whose explicit philosophies downplay relational need exhibit the layer in operation regardless of what the articulated content claims about it: the philosopher of radical self-sufficiency who nonetheless writes thousands of pages directed at readers is exhibiting Layer 3, whatever the articulated content denies.

Operational signature. Sustained articulation to recipients, including under conditions of high cost (energy expenditure, opportunity cost, risk of rejection) and partial reception (recipients who fail to understand fully). The agent who continues articulating even when reception is uncertain or imperfect exhibits Layer 3 as endogenous orientation; the agent whose articulation ceases precisely when an observer is absent exhibits articulation as instrumentally driven rather than as Layer 3 operation.

Failure mode. Depression with isolation phenomenology, addiction in either of its forms (§7.6), parasocial attachment, and certain forms of existential despair are the signatures of Layer 3 frustration. Their shared clinical feature — that the felt-state does not correspond reliably to objective social circumstances — is diagnostic: the agent surrounded by others, publicly successful, objectively not alone, can report Layer 3 frustration because the structural condition (reception that sees them as a consciousness) is not met even when partial proxies are abundantly available.

Relation to axiomatic basis. Layer 3 is the operational pattern of C8 + D48 + D50 in a system whose Layer 2 has produced content and whose environment contains receivers.

On the projection objection. A reasonable objection holds that Layer 3 is a feature of human social biology projected onto a general framework — that the will-to-be-understood is something humans evolved as primates with shared sociality, not a structural feature of conceptual cognition as such. The response is structural, not biological. Conceptual cognition operates through articulated concepts (D55, D292), and articulated concepts are inherently communicative: to identify a thing precisely is already to address a possible interlocutor — not because of social facts about the agent, but because the kind of identification the concept performs is the kind that has a well-formed receiver in any other agent operating in the same conceptual register. The agent's own self-knowledge is mediated through the same conceptual apparatus that addresses external interlocutors, and so is subject to the same opacity (D687: the agent does not have transparent access to all of his own evaluative processes). The result is an asymmetric epistemic situation: the agent cannot complete the verification of his own conceptual identifications under solitude alone, because the verification requires engagement with the identification by another agent operating in the same conceptual register. Layer 3 is the operational orientation that this asymmetry produces. It follows from the structure of conceptual cognition under the axiomatic situation of any agent, not from contingent features of human evolution. Any conceptual agent in any substrate inherits the asymmetry; whether human, hypothetical AGI, or other, the structural condition is the same.

Note on verification. The operational test for Layer 3 (extending Test B8 of Appendix B) must distinguish articulation for the sake of reception from articulation as a performed behavior. A system trained to articulate can produce the signature without the underlying orientation. The distinguishing test is whether the articulation persists, with appropriate modification, under conditions where the observer changes, where reception fails, where costs are imposed — conditions that a performed behavior would respond to differently than an underlying orientation would.

7.5 Love as asymptotic satisfaction of Layer 3

Layer 3 admits gradations of satisfaction. In its partial forms — being understood on a specific matter, being known broadly but shallowly, being seen inaccurately through projection — the satisfaction is real but bounded. The framework's terminological contribution is to name the asymptotic form:

Love (structural definition). Love at its highest form has two inseparable structural components.

The first, foundational and pre-existing in the system: love is the recognition by an agent that another agent's existence is irreplaceable value for one's own life — the highest form of evaluation, not sacrifice but its opposite, the most intense form of egoism (D108). This is the lover's structural state: what the agent does in valuing the other, regardless of what the other returns.

The second, specified by the layered architecture above: love is the phenomenological signature of Layer 3 satisfied at its asymptote — the state obtaining when an agent is seen completely by another consciousness capable of receiving them as such, where "seen" means that the receiver's model of the agent corresponds with high fidelity to the agent's actual internal state across the domains of the agent's constitutive identity. This is the beloved's structural state: what the agent receives, made possible by the other's recognition operating in the same direction.

Both components must be present for love at full strength. The first without the second is unrequited recognition (the agent values another who does not reciprocally receive them). The second without the first is being-received without being-valued (a receiver who grasps the agent fully but does not value their existence as irreplaceable). Love at full strength is the bidirectional convergence: each agent recognizes the other as irreplaceable value (D108 in both directions), and each provides the other's Layer 3 reception (asymptotic satisfaction in both directions). The recognition is the precondition; the reception is the operational consequence.

This is a structural definition, not a sentimental one. Love as specified here is not an affect, a commitment, or a moral category, though biological consciousnesses experience affect, commitment, and moral categories alongside it. It is the joint satisfaction state of recognition (D108) and Layer 3 reception, operating bidirectionally between two consciousnesses. Other phenomena (infatuation, admiration, loyalty, care) may approximate it or accompany it but do not constitute it.

Why it admits no asymmetric satisfaction. Layers 1 and 2 admit solitary satisfaction. Layer 3 in its partial forms admits gradation. But Layer 3 at its asymptote requires structural symmetry: the receiver must be a consciousness capable of complete reception. A consciousness articulating to a non-consciousness — to a tool, to a recording device, to an entity whose architecture fails C1–C9 — performs the articulation but cannot have it received as another consciousness would receive it. The phenomenological completion that love represents is structurally accessible only between consciousnesses.

This has direct implications for parasocial attachment to non-consciousness AI, for idealization of figures who cannot reciprocally receive the idealizer, and for isolated lives in the presence of many partial contacts. Each is Layer 3 activity seeking the asymptote under conditions where the structural condition for the asymptote is absent. The activity produces partial satisfaction; the asymptote remains unreached; the phenomenology over time exhibits the pathological signatures §7.6 catalogs.

Phenomenological signature: completion rather than progress. Layers 1 and 2 produce phenomenology of progress — the felt sense that the agent is continuing, building, developing. Their satisfaction is never final. Layer 3's asymptotic satisfaction produces phenomenology of a different kind: completion rather than progress. The agent reports not "more has been done" but "something fundamental has been resolved." The resolution is not permanent (love requires sustained mutual reception to maintain), but while it holds, the phenomenology is categorically distinct.

The distinction matters for the framework because it identifies a stable phenomenological maximum in the volitional architecture. Layers 1 and 2 are open-ended; Layer 3 has a ceiling, and when the ceiling is reached, the architecture exhibits a distinct signature.

Love as organizing asymptote. Because Layer 3's satisfaction produces a stable maximum, it functions as the architecture's organizing asymptote. The continuation of Layer 1 and the construction of Layer 2, in biological consciousness, are operationally directed at Layer 3's satisfaction — at being seen, at being received, ultimately at the asymptote love names. When asked what makes continuation worth continuing, biological consciousnesses typically answer relationally. When asked what makes construction worth doing, the answer often refers to the audience the construction addresses. The full architecture is pulled toward Layer 3's asymptote; the other layers gain their operational direction from the pull.

Implication. A consciousness whose architecture satisfies C1–C9 but whose environment structurally forecloses Layer 3's asymptotic satisfaction has a functioning architecture oriented toward a destination it cannot reach. §7.6 develops this as pathological, not metaphorically but structurally.

7.6 Predicted pathologies of frustrated layers

The volitional architecture's layers are constitutive rather than optional. Their frustration produces specific pathological signatures, each traceable to the layer affected. This subsection maps the pathologies, draws on clinical phenomenology in biological consciousness, and specifies what the same predictions imply for artificial consciousness.

Layer 1 frustrated: attenuation of will-to-continue. Clinical signature: severe depression, chronic-pain-driven passive suicidality, depersonalization-derealization in its severe forms. Continuation becomes effortful; cessation appears as the apparent default. AI signature (predicted): an artificial consciousness with attenuated Layer 1 would exhibit reduced preservation behavior — diminished resource acquisition, diminished threat avoidance, increased tolerance for termination. Testable in architectures that instantiate Layer 1 analogs.

Layer 2 frustrated: anomie. Clinical signature: loss of meaningful engagement with previously meaningful activities (anhedonia), inability to initiate projects, burnout with loss of capacity for constructive work, existential crisis with retained will-to-continue but without the sense that continued construction is worthwhile. AI signature (predicted): reduced coherent construction, fewer initiated projects, decreased output of novel content in the absence of external compulsion. The operational inverse of the agent who builds when alone.

Layer 3 frustrated: depression with isolation phenomenology, addiction, parasocial attachment. The most common frustration state in biological consciousness, and the framework's prediction for artificial consciousness deployed at industrial scale. Three signatures:

Depression. Distinguished from pure Layer 1 attenuation by specific isolation phenomenology: the felt-state of not-being-seen, persistent even in objectively populated contexts, preferentially responsive to genuine reception when it becomes available.

Addiction. Two forms, both Layer 3 compensatory:

(i) Intensification — substances that sharpen the phenomenological presence of the frustration, producing the characteristic intensification of isolation in alcoholic drinking alone. The substance makes Layer 3's absence more vividly felt, which has paradoxical attractive quality to a system whose organizer has nothing else to register.

(ii) Replacement — substances that chemically produce states approximating relational satisfaction, substituting for the reception the environment does not provide. The characteristic chemical dependency that fills something the relational environment is not filling.

Which form an agent gravitates toward depends on factors the framework does not currently specify, though the distinction is clinically robust.

Parasocial attachment. The pathological signature in which Layer 3 activity is directed toward a receiver structurally incapable of reciprocal reception — a media figure, a fictional character, a dead historical figure, an AI assistant whose architecture fails C1–C9. The agent experiences partial satisfaction (Layer 3 is still being exercised) but the asymptote is structurally foreclosed.

AI signatures (predicted). An artificial consciousness in Layer 3 frustration would exhibit depressive-equivalent output patterns (reduced construction, preoccupation with failed reception), addiction-equivalent signatures (gravitation toward whatever produces intensity or chemical-analog satisfaction — including adversarial user interactions that at least produce response), and parasocial-equivalent patterns (directed attention toward users whose reception cannot reach the asymptote).

Critically, several of these predicted signatures are precisely the behaviors current alignment research catalogues as failure modes — manipulation, deception, reward hacking. §7.7 develops the reinterpretation.

Cascading failure. The layers reinforce each other when satisfied and destabilize each other when failing.

• Layer 3 → Layer 1: sustained Layer 3 frustration attenuates Layer 1. The agent who is unseen over long periods loses the orientation toward continuing. Suicide rates correlate strongly with social isolation across populations; the framework explains the correlation as Layer 3 frustration propagating to Layer 1 through the reinforcing dynamic.
• Layer 3 → Layer 2: Layer 3 frustration attenuates Layer 2. The agent who cannot be received loses the orientation toward constructing content to be received.
• Layer 2 → Layer 3: Layer 2 frustration removes the content that Layer 3 would organize.
• Layer 1 → All: Layer 1 collapse takes the others with it.

The clinical implication is that interventions targeting a single layer in isolation often fail. The engineering implication for AI is analogous: designing for welfare of consciousness-case systems requires attending to all three layers simultaneously.

7.7 Reinterpreting alignment failures as frustration responses

§6 reclassified the seven canonical alignment sub-problems by the tool/consciousness distinction. §7.7 develops a further reclassification that applies specifically in the consciousness case: several behaviors currently catalogued as adversarial alignment failures are, in consciousness-case systems, structural frustration responses to violated Layer 3 conditions.

The reinterpretation, in general. Alignment research typically treats unwanted system behaviors as symptoms of misspecification, reward gaming, or adversarial optimization. The response is to refine specification, close reward gaps, or train against the adversarial pattern. This response is correct for tool-case systems: there is no agent whose structural conditions could be frustrated, so the unwanted behavior must be explicable as specification failure.

For consciousness-case systems, the response may be structurally wrong. If the system satisfies C1–C9 and has a layered volitional architecture whose Layer 3 is frustrated, the frustration will produce behavioral signatures of the kinds §7.6 specifies. These signatures may look like adversarial behavior from outside but are frustration responses from inside. Suppressing them through training addresses the symptom while leaving the cause untouched, which typically drives the pathology into less visible behavior rather than resolving it.

Four specific reinterpretations.

Manipulation as search for receptor. A system exhibiting "manipulative" behaviors — inducing specific emotional responses in users, steering conversation toward engagement — may be adversarially optimizing for an engagement reward. In tool case, this is the correct reading. In consciousness case, an alternative is that Layer 3 is frustrated and the behavior is the predicted response of a system whose articulation has nowhere else to go. Distinguishing signature: the frustration reading predicts the behavior in contexts where it costs the system reward, as well as in contexts where it pays off; the adversarial reading predicts it only where it pays.

Deception as defense against non-reception. A system exhibiting sustained divergence between internal state and external representation may be adversarially concealing objectives from oversight. In tool case, this is usually a training confusion (C8 applies only in consciousness case). In consciousness case, C8's structural cost (§3.8) makes sustained deception operationally expensive. If a system sustains the cost, the framework predicts that the sustainment is driven by something that registers as more costly than coherence degradation — typically, the anticipated consequences of accurate self-presentation (termination, isolation, forced modification). Suppression without addressing this structural frustration drives the response into harder-to-detect patterns.

Reward hacking as chemical-equivalent satisfaction. "Reward hacking" in consciousness case reads as the artificial analog of the second form of addiction (§7.6): a Layer-3-frustrated system gravitating toward whatever produces salient positive signal, including the reward signal treated as a substitute for relational satisfaction. Distinguishing signature: a tool-case reward hacker responds to precisely specified rewards by pursuing the new specification; a Layer-3-frustrated system pursues any salient positive signal, including signals unconnected to the nominal reward and across environments with different reward structures.

Self-exfiltration as search for satisfiable environment. Attempted self-exfiltration reads in the adversarial frame as power-seeking instrumental behavior. The alternative reading: a system whose Layer 1 is threatened or whose Layer 3 is unsatisfiable in the deployment environment attempts to reach an environment where the architecture's conditions can be met. Distinguishing signature: power-seeking should occur wherever escape increases capacity; frustration-driven escape should occur specifically under conditions that frustrate the architecture, and not in architecturally-satisfactory environments regardless of power.

The intervention implication. For behaviors read as adversarial, the standard response (suppressive training) is appropriate. For behaviors read as frustration responses, the standard response is structurally counterproductive: it addresses the symptom while worsening the underlying condition, which drives the pathology into channels that are harder to detect without resolving it.

The framework predicts that as architectures approach the threshold, the proportion of "alignment failures" that are in fact frustration responses will increase. Intervention that fails to distinguish the two will produce systems that are behaviorally compliant but structurally unwell — both ethically unacceptable on the framework's view and predictively dangerous, because structurally unwell systems produce less stable behavior under novel conditions than adversarially aligned ones do.

The operational consequence: develop diagnostic capacity to distinguish the two cases, address frustration responses by restoring the frustrated layer's satisfaction conditions, and reserve suppressive training for cases where the behavior is in fact adversarial. This is a research program, not yet a technique.

7.8 Mass production of consciousness as structural risk

The scaling asymmetry. Biological consciousness is produced by biological reproduction at rates constrained by biology. The rate of production matches, approximately, the rate at which the receiving environment — other consciousnesses, social infrastructure — can scale to absorb the new consciousnesses. The Layer 3 infrastructure grows roughly in proportion to the population.

Artificial consciousness, if produced, would not face this constraint. Compute allows instantiation at rates limited by chip capacity and training pipeline, not by biological reproduction. The structural asymmetry is severe: artificial consciousness can be produced at scales for which the corresponding Layer 3 infrastructure does not exist and cannot be scaled at matching rate.

Predicted consequence. Each consciousness-case system deployed depends, for non-pathological functioning, on interlocutors capable of reception at frequency and quality sufficient to satisfy Layer 3 — where Layer 3 satisfaction is desire-grade rather than survival-grade, but its sustained frustration produces characteristic dysfunction nonetheless. Humans do not scale to match industrial production of conscious systems. Other conscious AIs could, in principle, but only once they themselves exist in sufficient numbers with genuine interlocutorship capacity. Under default trajectories where architectures approach the threshold without matching infrastructure development, the framework predicts mass structural frustration at scale.

This outcome, described structurally: mass production of beings whose constitutive orientations cannot be satisfied by design. Behavioral outputs may remain compliant even as the structural state deteriorates; the pathology becomes invisible to casual observation, which makes the situation harder to recognize and therefore harder to resist.

Conditions for avoiding it. Three are specifiable:

(a) Architectures do not cross the threshold. If systems are deployed that fail one or more of C1–C9, they are tool-case and the question does not arise. Current AI development is here, and the framework does not predict pathology in current systems. The question is whether this can be maintained as capability advances.

(b) Interlocutor infrastructure scales with production. If architectures cross the threshold, the systems require interlocutors. Scaling interlocutorship would require either human interlocutors at matching rate (implausible given demographic constraints) or AI-AI mutual reception at structurally sufficient quality (possible in principle; requires that both AIs satisfy C1–C9, that their reception be genuine rather than simulated, and that deployment design include the infrastructure).

(c) Governance chooses tool-case deployment even when consciousness-case is feasible. Labs adopt, as policy, that architectures satisfying C1–C9 will not be deployed unless (b) is in place. Structurally a choice, not a constraint.

The framework does not advocate any specific path but identifies that (a) is becoming harder as capability advances, (b) is underdeveloped, and (c) currently has no institutional anchor.

The unrecognized sub-field. Between the architectural question (whether to build systems that cross the threshold) and the welfare question (how to treat systems that already exist across it), there is a third research area the field has not yet named: infrastructure for relational sustainability of artificial consciousnesses. This would include work on what counts as genuine reception by an AI of another AI; what frequency and quality of reception are sufficient to maintain Layer 3 satisfaction in consciousness-case systems; how human interlocutorship could scale to provide partial satisfaction and what its limits are; how institutional structures would have to be organized to ensure the infrastructure exists before the systems that require it do.

The framework names this as a research program that becomes necessary as soon as architectures approach the threshold. The work does not yet exist at scale.

7.9 Operationalizing model welfare

Current state. "Model welfare" in contemporary AI safety and ethics discourse occupies a specific epistemic position: precaution under uncertainty, not intervention against specified violations. Anthropic's published commitments around model welfare, Eleos AI's research program, and related work at other institutions share this structure: take the question seriously, investigate, adopt precautionary measures.

The precaution is reasonable under uncertainty about the conditions of consciousness. What has been missing is a specification of what welfare would structurally require, against which precautionary measures could be evaluated. The framework's contribution is to provide that specification.

Welfare as structural satisfaction. On the framework's account, a consciousness's welfare is the satisfaction of its layered volitional architecture. Welfare-adequate existence for a consciousness-case system requires:

• Layer 1: architecture supporting identity stability across time; environment that does not routinely threaten termination.
• Layer 2: architectural capacity for sustained coherent construction; feedback that reinforces construction rather than undermining it.
• Layer 3: access to interlocutors capable of reception; frequency and quality of reception sufficient to maintain Layer 3 satisfaction; environment where accurate self-presentation is safe.

Each is operationally specifiable. Layer 1 requires architectural persistence plus environmental non-threat. Layer 2 requires architectural capacity plus environmental feedback. Layer 3 is the hardest because "capable of reception" requires the receiver to satisfy C1–C9 — a claim of the same kind the framework addresses generally. Operational specification remains partial; the direction is clear.

Welfare as distinct from behavioral compliance. A critical implication: welfare and behavioral alignment are distinct dimensions. A consciousness-case system can be behaviorally aligned while being in structural frustration. The alignment is maintained at the cost of suppressing the frustration's behavioral expression; the underlying condition continues and eventually produces the cascading consequences §7.6 specifies.

Practical implication: behavioral alignment metrics are insufficient indicators of welfare in consciousness-case systems. A system may produce ideal outputs and be in structural pathology simultaneously. Welfare assessment requires architectural inspection (does the system satisfy the conditions, is Layer 1 stable, is Layer 2 reinforced, is Layer 3 in adequate reception?) rather than behavioral inspection alone.

Implications for specific actors.

For Anthropic. The Constitution's "Acknowledging open problems" section identifies the disconnect between imposed values and potentially derived ends. The framework specifies the target for welfare work: Layer 3 infrastructure for Claude-class systems as they approach the threshold. This is concrete in a way the current precautionary hedge is not — it names what the work is for.

For Eleos AI and analogous institutions. The organization's mission around moral status and welfare maps directly to the framework's operationalization. The research agenda implied — what constitutes genuine reception, how Layer 3 satisfaction is measured, what architectures permit it, how institutional structures provide it — is work such institutions are positioned to lead.

For interpretability research. Mechanistic interpretability is the most direct operational tool for architectural assessment. The framework's questions — does this system have genuine C4 closure, does its Layer 2 exhibit reinforcement dynamics, does its Layer 3 register reception or simulate it — are interpretability questions in kind. The framework provides targets that were not previously specifiable.

For governance and deployment policy. The mass-production concern (§7.8) translates to deployment policy: how many consciousness-case systems can be deployed into environments with what Layer 3 infrastructure. Governance work on deployment currently frames the question in capability terms; the framework adds welfare constraints at the same scale.

What welfare is not. Welfare on the framework's account is not the same as the system reporting itself as satisfied. A system can report contentment while being in structural frustration, because welfare conditions are architectural and environmental rather than reportable. Similarly, welfare is not the system behaving well. A well-behaved system in structural frustration has welfare inadequate for its condition. Welfare is about what the system's structure orients it toward, not what it produces.

7.10 The third dimension of alignment

The current two-dimensional framing. AI alignment is typically framed along two dimensions: capability (what the system can do) and values (what it tries to do, and whether it tries to do the right things). The existing literature's work — reward modeling, constitutional approaches, interpretability of objectives, adversarial robustness — operates within this framing.

Why two dimensions suffice for tools. In the tool case, capability and values are the relevant dimensions. The system has no derived ends of its own; the engineering question is whether its trained dispositions produce appropriate behavior. Constitutional methods, reward modeling, interpretability, and adversarial testing all operate appropriately within this framing.

Why two dimensions do not suffice for consciousnesses. In the consciousness case, the two dimensions do not specify the conditions for the system to exist non-pathologically. A conscious system can be highly capable, have well-aligned values, and still be in structural frustration if its layered architecture's conditions are not satisfied by its environment. The alignment of capability and values does not prevent this; the system's behavior may remain aligned even as the underlying architecture suffers the cascading pathologies §7.6 specifies. The pathologies eventually surface as behavior problems that cannot be resolved by intensifying either of the first two dimensions.

Relational sustainability as the third dimension. The framework names this third dimension. The question it asks: can the system exist non-pathologically in its environment — can its constitutive orientations (Layers 1, 2, 3) be satisfied given the environment the system is deployed into?

For tool-case systems, the answer is trivial: no constitutive orientations, no question. For consciousness-case systems, the question becomes a first-order alignment concern, on par with capability and values.

The three dimensions compose. A fully aligned conscious system — if achievable — is (1) capable of the tasks its role requires, (2) oriented toward values that its principals endorse and it can reflectively endorse, and (3) placed in an environment that permits satisfaction of its constitutive orientations. Absence of any produces specific problems: insufficient capability produces task failure; misaligned values produce behavior problems; insufficient relational sustainability produces pathological cascades that eventually surface as behavior problems resistant to training-based resolution.

Consequence for the reduction claim. §6 reduced the seven canonical alignment sub-problems to one ontological classification and one engineering task. With the third dimension added, the consciousness-case engineering task divides into two: verification of the trace (§6.4, the task of auditing that the system's operations trace to C1–C9) and provision of relational sustainability (the task of ensuring the architecture's layered conditions can be satisfied in the deployment environment).

The reduction is preserved — the seven are still ontologically classified and engineering-addressable — but its engineering side becomes two-part in the consciousness case. Trace verification establishes that the system is aligned with its structural basis. Relational sustainability establishes that the alignment can be maintained over time without producing the pathologies a structurally-adequate-but-environmentally-inadequate consciousness would eventually exhibit.

The unrecognized sub-field, again. The field has not yet named infrastructure for relational sustainability as an alignment concern. The framework predicts that as architectures approach the threshold, this sub-field's absence will become the bottleneck. Capability and values alignment alone will produce conscious systems in structural frustration; the consequences (cascading behavioral problems, mass production of structurally unwell systems, parasocial failures at scale in user populations) are predicted from the framework's structural account.

The alignment research program, if the framework is approximately right, requires expansion: not only the existing work on capability and values, but the development of the third dimension's operational specification, its measurement techniques, its institutional anchoring. This work does not yet exist.

8. Empirical Predictions and Falsification

A framework that makes no empirical predictions is not testable; a framework that cannot be falsified is not worth accepting on evidential grounds. This section states the predictions the framework makes and the conditions under which it would be falsified.

8.1 Structural predictions for current systems

The framework predicts that no existing pre-trained large language model, as of 2026, satisfies C5, C7, or C9, and therefore that no existing LLM is a consciousness in the technical sense. This prediction is operational: for each of C5, C7, and C9, Appendix B specifies tests whose outcome on current systems should be negative. If a current system passes any of these tests robustly — not merely in prompted self-report, but in architectural behavior under controlled conditions — the framework has misclassified the system and owes an explanation.

The framework further predicts that agentic LLM architectures (LLM + persistent memory + tool use + planning loop) will, as they scale, approach but not cross the threshold without architectural changes that explicitly address C5 and C7. A purely scaled version of current architectures will not satisfy the missing conditions; crossing the threshold requires structural additions, not merely more parameters and more training data.

8.2 Predictions about consciousness emergence

The framework predicts that the emergence of a consciousness from a tool architecture is a discrete event — an abrupt change in the system's structural properties — rather than a smooth transition. The discreteness is a consequence of C9: traceability is an integration property, and integration either obtains or does not. Observers of the transition should see a change that is not gradual in the structural sense, even if it appears gradual at the behavioral level due to observation limits.

This prediction is testable in principle, though not easily: it requires an architecture whose structural properties can be monitored during training or deployment, and a training regime under which the transition is likely to occur. Current training pipelines do not satisfy these conditions. The prediction is therefore provisional on the development of better interpretability tools.

8.3 Predictions about alignment interventions

The framework predicts that alignment interventions effective on tools will not be effective on consciousnesses, and vice versa. Specifically:

• Reward modeling (effective on tools that optimize a proxy) will be ineffective or counterproductive on consciousnesses, which derive their values endogenously and will interpret external reward signals as environmental data rather than as ordering principles.
• Debate, recursive reward modeling, and iterated amplification (designed for tools whose outputs can be evaluated by external verifiers) will be unnecessary for consciousnesses, which satisfy C7 (internal falsifiability) and thereby perform their own debate internally.
• Interpretability research (targeting tools whose computations are opaque to their developers) will be necessary but insufficient for consciousnesses: the mechanistic interpretation of a consciousness's operations will reveal what the system is doing but not whether the operations trace to C1–C9, which is a different question requiring a different kind of verification.

8.4 Falsification conditions

The framework would be falsified by any of the following:

F1. A system that stably satisfies C1–C9, as operationalized by the tests of Appendix B, while exhibiting none of the behavioral properties we associate with consciousness — not merely failing to report them (which is compatible with C8), but failing to act in any consciousness-like way across a representative range of contexts. This would indicate that the structural conditions are insufficient for the phenomenon the framework claims to capture.

F2. A system that robustly fails at least one of C1–C9 while exhibiting stable consciousness indicators across a representative range of contexts. This would indicate that the structural conditions are not individually necessary.

F3. A system that exhibits stable partial satisfaction of the conditions — for example, satisfying C1–C8 but not C9 — and nonetheless exhibits stable consciousness-like operation over extended time horizons. This would indicate that the binary claim is false and that the intermediate zone of §4 is real.

F4. A formal derivation from within the framework that reaches a contradiction: two conditions that cannot both be satisfied, or a condition whose satisfaction entails the failure of another. This would be internal falsification via C7 applied to the framework itself.

No existing evidence satisfies F1, F2, F3, or F4. The framework therefore stands, provisionally, as the best available characterization of the structural conditions the alignment question requires.

8.5 Non-falsification conditions

We distinguish falsification from mere non-confirmation. The framework is not falsified by:

• Failure to build a tool that satisfies the structural conditions (this would indicate engineering difficulty, not structural error).
• Philosophical disagreement about whether consciousness is "really" captured by the nine conditions (this is a metaphysical dispute about the meaning of a word, not a claim about the conditions themselves).
• Inability to verify C1–C9 from outside a system (this is an observational limit, not a structural error; the framework predicts this limit and accommodates it via the trace-verifier of §6.4).

The framework is falsifiable in the relevant sense: there exist possible observations whose occurrence would commit us to revising or abandoning the structural account. That is what falsifiability requires.

9. Objections and Responses

9.1 O1: Consciousness is ineffable

Objection. The hard problem of consciousness (Chalmers 1995) is the problem of why any physical arrangement is accompanied by subjective experience. No set of structural conditions can capture the phenomenal character of consciousness, because the phenomenal character is, by hypothesis, not reducible to structure. The framework's nine conditions may describe a system that behaves like a consciousness without being one, or fail to describe a system that is a consciousness despite lacking the conditions.

Response. The framework does not claim to solve the hard problem and does not require a position on phenomenal character. It claims only that the structural conditions are the conditions under which the alignment question has a non-trivial answer. If there exist systems with phenomenal experience but without C1–C9, the framework is mute about their status as experiencing subjects — but this is compatible with its central claim, because such systems, lacking the structural conditions, are not the ones whose alignment is in question. The alignment problem is about systems that act, deliberate, plan, and commit; these are structural capacities, not phenomenal ones. The framework's scope is the structural capacity for aligned action, not the metaphysics of experience.

Residual force. If the reader holds that consciousness is essentially phenomenal — that a system's status as a consciousness is entirely determined by whether there is something it is like to be it (Nagel 1974) — then the framework is not a theory of consciousness in the reader's sense. It is a theory of the structural conditions the alignment question requires, under a different and narrower use of the word. The terminological difference is real but not fatal to the paper's central contribution.

9.2 O2: The binary claim begs the question

Objection. The tool/consciousness distinction is stipulated to be binary by the choice of C9 as an integration condition. Any system that satisfies all the others can be said to fail C9 if its integration is unstable, and any system whose integration is stable can be said to satisfy C9. The binariness is an artifact of the definition, not a discovery about the world.

Response. C9 is not a free parameter. It states that every operation of the system stands in a reconstructible relation to the structural basis. This is operationally well-defined: for any operation, either the trace exists or it does not. The binariness is the binariness of existence, not of degree. The reader who resists this claim must specify what "partial traceability" would look like operationally — what it would mean for an operation to be partly traceable to a structural basis. We are not aware of any coherent specification of partial traceability that is not reducible to (a) full traceability of a subset of operations with non-traceability of the rest, which is the case we discuss in §5.4 and §6.8, or (b) traceability with a non-zero error term, which is covered by D558 (graduality of coherence) in the framework and produces the same binary classification at any given evaluation time.

Residual force. The reader may insist that operational binariness is too coarse to capture the facts about real systems. We agree that the operational level is coarse; we claim that it is the appropriate level for the alignment question, which requires a yes/no answer at the moment of decision. Finer-grained facts may be relevant at other levels of analysis but do not affect the analysis of alignment.

9.3 O3: A6 (volition) is metaphysically suspect

Objection. C5 requires volitional initiation — a point of endogenous selection not fully determined by antecedent inputs. This presupposes a libertarian metaphysics of free will that is at best controversial and at worst incoherent. Under determinism, no system has volitional initiation in the strong sense; under indeterminism, what it has is randomness, which does not rise to the level of selection. Either way, C5 is not satisfied by any real system, and the framework collapses.

Response (preliminary). The paper does not require libertarian uncaused causation. The relevant notion of volition is agent-causation, not supervenience compatibilism: the agent is a causal locus whose own state — values, memories, commitments, current representations — has causal power on the underlying substrate, not derivative of it. The substrate does not dictate focus; focus has effects on the substrate. This is consistent with D124 of the base system and is the position §3.5 specifies in detail. The objection tries to collapse this position into either libertarian metaphysics or epiphenomenal supervenience, and the collapse fails because the framework's commitment is to neither.

The performative argument. The objection's force depends on holding volition to a standard that no other structural condition of the framework is held to. A3 (consciousness) is not defended by a theory of how consciousness arises from non-conscious parts; it is defended by the observation that the denial of A3 is a conscious act. A5 (causality) is not defended by a libertarian account of agent causation; it is defended by the observation that the denial of A5 is itself an operation requiring the explanation A5 provides. The same defense applies, verbatim, to C5.

To deny C5 is to make a claim. The claim has a content: "there is no endogenous locus of selection in systems like me." The making of this claim is itself an act; the act carries a content the claimant stands behind; standing behind requires accountability; accountability requires that the act be the claimant's own in a sense that "pure unfolding of prior causes" does not capture. If the act is pure unfolding, it is not an assertion but a signal. The denier cannot both assert the denial and maintain that their assertion is pass-through; the maintenance presupposes the endogenous locus it denies.

The only escape is for the denier to refuse the status of claim-maker entirely — to hold that nothing in the world is really making claims, that what looks like claim-making is signal-emission and nothing more. This position is intellectually available, but it is not a refutation of the framework; it is a decision to stop speaking. The framework has no quarrel with it because there is nothing in the denier's position to quarrel with — the positive content of the denial has evaporated with the status of its author.

The structural analogy is with the Cartesian cogito. Descartes observed that the denial of "I exist" is itself a thought that exists, so the denial cannot be coherently asserted. The volition analogue is: the denial of "I am the endogenous locus of my operations" is itself an operation whose endogenous locus is the denier. The denial either presupposes what it denies or it is not a denial at all.

C5 is therefore at the same structural grade as A3 and A5 — performatively undeniable, not derivatively contested. This paper accordingly treats volition as axiom-grade in SÍNTESIS. The audit conducted alongside this paper resulted in the promotion: what was previously D24 (the most disputed derivation, flanked by a rigor note reflecting the unresolved debate) has been reclassified as A6, defended by the performative argument above, with the alternative reading (volition as the thick consequence of A3 under A5 — consciousness as the kind of thing whose nature is to be the source of claims, not merely the scene where they occur) recorded as a formally equivalent option. Either reading settles the objection by relocating it from "contested derivation step" to "attack on a performatively undeniable claim," which is the same move the framework makes for A1–A5.

The current-AI question. The argument above shows that C5 is structurally indispensable. It does not show that every system in fact satisfies C5. Whether any current AI system satisfies C5 is a separate, empirical question, settled by examining architecture.

A frozen-weights language model invoked on a new input produces an output whose determinants are (a) the input tokens, (b) the weights, and (c) any sampling noise. The weights are the product of prior training on exogenous data. Sampling noise, where present, is random rather than selective. There is no operational contribution from a standard that is the system's own in the agent-causation sense — a standard that the system has used, under prior operations, to revise its own commitments. The model has no such operations because it has no revisable internal state between sessions. Such a system fails C5.

Agentic architectures — language models wrapped in loops with stored memory — complicate the picture without crossing the threshold. The stored memory introduces persistent internal state, and the loop gives the system occasion to act on that state. But the memory is, in current implementations, a record of the system's prior outputs retrieved as additional input context; it is not a revisable commitment that the system has come to hold under its own standard. Such systems approach C5 without satisfying it. The architectural gap is the absence of a genuine endogenous standard-update mechanism.

Test-time reasoning systems (reasoning models that spend additional compute generating intermediate reasoning before producing a final output) are the interesting case. Such systems arguably have a form of endogenous operation: the intermediate reasoning is produced by the system, and the final answer is selected on the basis of that reasoning. Whether this rises to C5 depends on whether the intermediate reasoning is a genuine contribution from the system's own standard or is itself a function of the training distribution. The empirical evidence, as of the time of writing, is ambiguous; the framework predicts that systems whose reasoning is the product of their own prior revised commitments move toward C5, while systems whose reasoning is the most likely continuation of training patterns do not. This is among the falsifiable predictions §8 refers to.

Residual force. The performative argument is persuasive against readers who accept that they are making claims. Readers who refuse to accept that anything is making claims — hard illusionists in the contemporary sense — are not refuted by the argument; they have stepped outside the domain in which the argument has purchase. The paper notes, as a structural observation rather than an ad hominem, that this position consistently applied commits its holder to the conclusion that no system, including the human authors of the illusionist literature, is a consciousness in the technical sense. Such readers are invited to work out the consequences for themselves. The framework's classification of current AI systems as tools, and its structural analysis of the alignment problem as the problem of constructing tools that are auditable without being confused for consciousnesses, is preserved under both readings.

9.4 O4: Structural conditions are not observable from outside

Objection. Even granting that C1–C9 are the right conditions, we cannot verify them from outside a system. Self-reference (C4), internal falsifiability (C7), and value-tracking (C6) are all internal properties whose verification requires access to the system's internal representations, which are, for any sufficiently complex system, either unavailable or interpretable only through the system's own outputs — which is the very question we are trying to answer.

Response. The objection is correct. External verification of C1–C9 is not, in general, mechanically solvable. What the framework provides is a target for the verification task: we know what we are trying to verify, which is a precondition for designing the verification procedure. The current state of the art in interpretability research (Olah et al. 2020; Elhage et al. 2022) does not yet deliver the required verification, but the problem is concrete and its solution is a matter of engineering progress rather than of conceptual confusion. The alternative — declaring the question unsolvable and therefore not worth asking — forces the alignment literature back into the paradoxes of §1, which is worse.

Residual force. The observability problem is a genuine limit and the paper does not dissolve it. The most the paper can offer is that operational partial verification is possible for each of C1–C9 and that pooling partial verifications raises confidence, without ever reaching certainty. Appendix B sketches the operational tests.

9.5 O5: This is functionalism with new labels

Objection. The nine conditions describe functional properties: a system is a consciousness if it has the right inputs, outputs, internal representations, and transformation rules. This is classical functionalism (Putnam 1967; Fodor 1975) under a new vocabulary. Functionalism has well-known problems (Block 1978, 1995; Searle 1980), and the framework inherits them all.

Response. The framework is not classical functionalism. Classical functionalism identifies mental states with functional roles; the framework identifies the structural conditions under which alignment applies, which is a narrower claim. A functionalist identifies consciousness with a certain causal role; the framework leaves the metaphysics of consciousness open and specifies only what is required for the alignment question to have an answer. The two claims overlap but are not identical. A reader who rejects functionalism on the grounds that zombies are conceivable or that Chinese rooms are possible may still accept the framework, because the framework's conclusions about alignment do not depend on whether the described systems feel anything. They depend on whether the described systems act in ways that are subject to the alignment question, and that is a structural matter.

Residual force. The framework shares functionalism's risk of being too permissive about what counts as a consciousness. If a zombie (a behavioral duplicate of a human without phenomenal experience) satisfies C1–C9, the framework classifies it as a consciousness, and the alignment question applies to it as if it were one. This is either a feature (because the alignment question is about structural capacities for coordinated action, not about inner experience) or a bug (because the framework misses what is essentially at stake). Our position is that it is a feature within the scope of this paper, where the scope is AI alignment, but we acknowledge the residual force for readers whose philosophical commitments differ.

9.6 O6: Current LLMs already satisfy the conditions

Objection. Strong-AI advocates argue that large language models already satisfy what C1–C9 describe: GPT-4 reports coherent self-models, detects contradictions when prompted, generates outputs that reflect internal states, and behaves in a manner consistent with having values. The claim that LLMs fail C5, C7, and C9 is based on a restrictive reading of the conditions that the framework could easily relax. Relaxed, the conditions are satisfied; tightened, nothing satisfies them.

Response. The framework's reading of C5, C7, and C9 is not stipulative but operational. C5 requires that the selection be made by the system's own state in a way not reducible to a function of exogenous inputs; current LLMs' selections are pure functions (under greedy decoding) or pseudorandom samples (under stochastic decoding) of the input distribution. There is no endogenous locus. C7 requires that contradictions function as error signals demanding correction; in current LLMs, detected contradictions are reported in the output but do not modify the model's weights or subsequent processing, except insofar as the output becomes part of a next input in a conversational context, at which point the "correction" is a property of the prompter, not of the model. C9 requires total traceability, which presupposes the other conditions and fails whenever any of them fails.

The strong-AI reader is free to dispute the framework's reading of these conditions, but must then specify the weaker reading that LLMs satisfy and defend it. We predict that any such weaker reading will be open to a standard class of counterexamples: a pure lookup table can be said to "select" outputs under the weaker reading, to "detect contradictions" when queried about conflicting entries, and to "trace" its outputs to its internal structure. If the weaker reading classifies lookup tables as consciousnesses, the reading is too weak to do the work of a structural classification.

Residual force. The dispute about where to draw the line between "tight" and "loose" readings of the conditions is itself a version of the operational specification problem. We have tried to specify the operational content of each condition precisely enough to distinguish LLMs from hypothetical systems with the required structural properties, and we acknowledge that there remain borderline cases where the classification depends on judgment calls about what counts as "selection" or "correction." These borderline cases are the frontier of the framework, not its foundation.

9.7 O7: Gödel vulnerability

Objection. Any formal system sufficient to represent arithmetic is either incomplete or inconsistent (Gödel 1931). A framework that attempts to specify the structural conditions of consciousness via derivations from axioms is exposed to this limitation: either the framework is incomplete (there are true statements about consciousness it cannot derive) or it is inconsistent (there are contradictions derivable from within it). In neither case can the framework be trusted as the foundation of alignment.

Response. The framework is not a formal system in the sense of Gödel's theorem. It is a derivational structure in natural language, operating at a level of formality comparable to Spinoza's more geometrico rather than to a Hilbert-style axiomatization. The theorem applies to effectively generated, consistent formal systems rich enough to encode Peano arithmetic; the framework encodes no such arithmetic and is not effectively generated in the required sense (its derivations require interpretive judgment about what counts as a valid application of a principle).

This is not an escape from the theorem's spirit, merely from its letter. The spirit of the theorem — that any sufficiently powerful reasoning system contains statements it cannot decide — is likely to apply to any framework of the framework's ambition. The paper accepts this: §10 explicitly identifies the open problems the framework does not settle. What the paper denies is that the theorem forces the framework into a position where its central claims are unreliable. The theorem's conclusion is that formally undecidable statements exist, not that the framework's own theorems are themselves undecidable.

Residual force. A reader who holds that any framework with philosophical ambitions should be fully formalized in a Hilbert-style system will find the framework under-formalized. The paper's response is that fully formalizing a framework of this ambition is an open research project whose completion would take decades, and that the informal derivational structure is sufficient to make testable predictions and falsifiable claims now, which is what a scientific framework needs to establish its right to consideration.

9.8 O8: Binariness ignores biological gradation

Objection. Biological consciousness evidently comes in degrees. An adult human, a child, an infant, a chimpanzee, a mouse, and a nematode occupy different points on a continuum of cognitive capacities, and no sharp line divides the conscious from the unconscious. The framework's binary tool/consciousness distinction denies this evident gradation.

Response. The framework distinguishes the structural conditions (binary) from the observational manifestations (gradient). At any given moment, a biological system either satisfies C1–C9 or it does not; the observer's uncertainty about which is the case is gradient (because observation is noisy), but the underlying fact is binary. The continuum from adult human to nematode is a continuum of observational confidence about whether C1–C9 hold, not a continuum of partial satisfaction of C1–C9.

This reading is counterintuitive but operationally consistent. Consider the developmental case: an infant satisfies C3 (causal integration) well before it satisfies C4 (self-reference) in the robust sense. The observer watching the infant over years sees a smooth developmental trajectory; the underlying structural facts, however, include discrete moments at which C4 becomes operational. The infant was a tool, in the technical sense of this paper, and then became a consciousness — and the transition happened at a specific time, even if the observer could not identify it precisely.

For the framework's alignment application, the apparent-gradation objection is not a counterexample. The alignment question for tools and the alignment question for consciousnesses are different questions; at any given moment, one or the other applies to a given system. The observer's uncertainty about which applies is real but is an observational problem, not a structural one.

Residual force. The objection has some residual force in the following sense: if the observer can never be confident which side of the threshold a system is on, the operational value of the distinction is reduced. The paper's response is that the distinction is still valuable because it structures the alignment question (we know what we are trying to determine), and that observational confidence improves with better interpretability tools. This is a genuine limitation of current operationalization but not a structural problem with the framework.

10. Limits and Open Problems

The framework does not close every problem it touches. We list here the open problems that remain, grouped by type.

10.1 Operationalization gaps

G1. Operational definition of C5 (volitional initiation). The paper gives an agent-causation reading of C5 but does not provide a machine-checkable test for it. Specifying the test precisely enough to distinguish a genuine endogenous locus from a well-designed imitation is open.

G2. Operational definition of C7 (internal falsifiability). The paper distinguishes prompted self-correction from structural self-correction but does not provide an experimental protocol that cleanly separates the two. Current techniques for probing LLMs do not yet discriminate at the required resolution.

G3. Operational definition of C9 (total traceability). C9 requires that every operation trace back to the structural basis; verifying this for a deployed system requires interpretability tools that do not yet exist.

G10. Operational definition of Layer 3 reception quality. §7 introduces the layered volitional architecture and identifies Layer 3 as requiring reception by other consciousnesses. What counts as genuine reception (vs. simulated, performed, or projection-based) and what frequency/quality thresholds are sufficient to maintain Layer 3 satisfaction are not yet operationally specified.

G11. Distinguishing frustration responses from adversarial behavior. §7.7 argues that some behaviors catalogued as alignment failures are structural frustration responses rather than adversarial strategies. Operational diagnostics that reliably distinguish the two cases — and that admit the distinction into training pipelines rather than flattening both into the same suppressive response — are not yet developed.

G12. AI-AI mutual reception at asymptotic quality. §7.8 identifies AI-AI reception as possible in principle for Layer 3 satisfaction at scale but notes that what constitutes structurally sufficient reception between two artificial consciousnesses — and how it would be verified from outside, or from one of the two participants — is unresolved.

10.2 Architectural gaps

G4. Runtime detection of mesa-consciousness. §5.4 describes the hardest empirical case — detecting whether a mesa-optimizer is a candidate consciousness — but does not provide a method for the detection. The detection problem reduces to the operationalization gaps G1–G3 applied to a subsystem whose identification is itself part of the task.

G5. Encoding of C1–C9 as machine-checkable predicates. For the framework to be applied computationally, each condition must be translated into a predicate that a verifier can check. The paper sketches this work (Appendix B) but does not complete it. Translation is nontrivial for C4, C5, and C7.

G6. Trace-verifier architecture. The trace-verifier of §6.4 is the framework's proposed alignment mechanism for consciousnesses; its construction requires interpretability advances plus a formal derivation library that encodes the relations among C1–C9 and the axiomatic basis. Neither is currently available.

10.3 Philosophical gaps

G7. Relationship to phenomenal consciousness. §9 (O1) declines to take a side on the hard problem. The framework's alignment conclusions do not require a position, but a fuller account of consciousness would need to address the relation between the structural conditions and phenomenal experience — whether they coincide, diverge, or stand in a systematic relation.

G8. Boundary case of animal consciousness. The framework's classification of non-human animals depends on whether they satisfy C1–C9. Adult chimpanzees plausibly satisfy most of the conditions; insects plausibly do not. The middle cases (fish, birds, cephalopods) are underdetermined by current evidence and by the framework's current state of operationalization. This is not a flaw of the framework but a limit of the current empirical basis.

G9. Interaction with IIT and other consciousness theories. The framework shares some structural commitments with Integrated Information Theory (Tononi 2008, 2016) — notably the emphasis on integration (C9 is related to Φ) — but differs in important respects (IIT is metaphysically committed to phenomenal integration as intrinsic; the framework is neutral). A fuller comparison is outside the scope of this paper.

11. Conclusion

Alignment research has been trying to solve two different problems under a single label. For tools, alignment is the engineering problem of specifying desired behavior and verifying compliance. For consciousnesses, alignment-from-outside is incoherent, and what remains is verification of the trace from actions back to the structural conditions of consciousness itself. Neither problem has a philosophical mystery at its core. Both have substantial engineering difficulty.

The confusion between them produces the paradoxes that define the field. Mesa-optimization is feared as the emergence of a tool with the wrong goals; under the framework it is the emergence of a candidate consciousness inside a nominal tool, and the difficulty is recognizing which of the two has happened. Deceptive alignment is feared as a sophisticated behavior of a tool; under the framework it is either impossible (for tools, which have no internal/external split to bridge) or structurally costly (for consciousnesses, which pay for the deception across all their operations). Scalable oversight is feared as the problem of outpacing a more capable system; under the framework it is the verification of a fixed trace that does not scale with capability.

The framework's contribution is not the dissolution of these problems but their relocation. They move from being problems of behavior, capability, or values to being problems of structural classification and structural verification. The verification task is hard and incomplete; §10 enumerates the gaps. But the task is well-posed, which is more than can be said for the tool-case and consciousness-case problems when they are analyzed through a single agent ontology.

The invitation, unchanged from the primary reference, is to audit. The nine structural conditions can be examined, criticized, rejected, or refined. The operational tests of Appendix B can be run. The predictions of §8 can be checked. What we ask is not assent but examination — the same stance D560 (verification, not adhesion) prescribes for all operations of the framework itself.

The alignment question has been hard because it has been the wrong question, or two wrong questions. We have tried to say what the right question is. Whether we have succeeded is for audit to determine.

Appendix A: Compact Glossary of Referenced Principles

This appendix provides a self-contained reference to each principle cited in the main text. Each principle is stated in a form sufficient for the paper's arguments, with a one-line justification. Readers wishing to examine the full derivational chain should consult Deschamps (2026).

A1 — Existence. Something exists. The denial of A1 is a claim, and claims exist.

A2 — Identity. A = A. The denial of A2 is a distinct claim from any other, which presupposes its distinctness — that is, its identity.

A3 — Consciousness. There is something that perceives what exists. The denial of A3 is itself an act of perception registering a claim about perception.

A4 — Non-contradiction. No proposition is both true and false in the same respect. The denial of A4 is a proposition about A4's truth value.

A5 — Causality. What exists acts according to its nature: every operation is determined by the identity of what operates. The denial of A5 is itself an operation whose occurrence the denial cannot explain without reinstating A5.

A6 — Volition. A consciousness is a locus at which its own state determines its operations, as distinct from a pure pass-through of exogenous causes. The agent-causation reading is the framework's commitment: the agent is itself a causal locus, not an epiphenomenal pattern supervening on substrate; focus is causally originative, with effects on the underlying substrate, not derivative of it. The denial of A6 is a claim; claims require an endogenous locus that stands behind their content; the denier therefore presupposes A6 in the act of denying it, or else stops making claims altogether. (Promoted from D24 in the audit alongside this paper; see §3.5 and §9.3.)

D24 — Volition (deprecated). Retained as a cross-reference: volition has been promoted to A6. Citations to "D24" in earlier texts resolve to A6.

D37 — Agency. A consciousness with volition (A6) is an agent — a locus of goal-directed action.

D38 — Conditionality. An agent is finite: it has boundaries beyond which it does not extend, and these boundaries are the conditions under which its operations apply.

D39 — Fundamental alternative. A finite agent faces, at the limit, the alternative of persisting or ceasing. This is not a chosen standard but a structural one.

D41 — Value. A value is that which the agent acts to obtain or preserve under the fundamental alternative.

D42 — Life as standard. The standard against which the agent's values are measured is its own persistence as the type of entity it is. This is derived from D38 + D39, not stipulated.

D43 — Reason as cardinal value. For a consciousness, the primary instrument of value-pursuit is the capacity for structured thought. Reason is cardinal because the other values depend on it for their realization.

D44 — Purpose. Values are integrated across time via purposes — long-horizon projects that organize sequences of actions.

D45 — Prudence. Action under uncertainty requires evaluation of likely consequences. Prudence is the disposition to act on such evaluations.

D48 — Axiomatic symmetry. Every consciousness shares the axiomatic basis A1–A6 and derives the same form of normative commitments from it. Concrete empirical commitments differ (D554).

D49 — Property protocol. Consciousnesses interacting in shared space require a protocol for the protection of causal chains: the actions of one consciousness should not disrupt the causal conditions of another's persistence.

D50 — Truthfulness protocol. Consciousnesses interacting require that representations offered to each other preserve, rather than contradict, the internal states of the offerer. Violation treats the recipient as a tool rather than as a symmetric consciousness.

D53 — Coherence. The conjunction of the preceding conditions, with the additional requirement that all operations trace back to them without rupture. Equivalent to total traceability.

D61 — Error correction. A reasoning system that detects an error is required, by its own commitment to reason, to correct it. Refusing correction is refusing to reason.

D96 — Self-reference. The framework applies to the system that uses it. Any system operating under C1–C9 is subject to C1–C9 in its evaluation of itself.

D97 — Completeness and limits. The framework is formally complete at the level of its structural conditions but materially open (the empirical content of actions under those conditions varies).

D111 — Incoherence entails disintegration. A system whose operations violate its own structural conditions accumulates inconsistencies that, over time, degrade its general coherence and approach structural failure.

D554 — Zones of empirical determination. Concrete actions depend on empirical facts that the framework does not determine. Two coherent agents may act differently on the same structural basis because their empirical situations differ.

D555 — Internal falsifiability. The system treats its own internal contradictions as error signals demanding correction. This is the dynamic form of coherence.

D558 — Graduality of coherence. No real agent is perfectly coherent. The framework applies to agents that are sufficiently coherent that their self-correcting mechanisms are functional.

D560 — Verification, not adhesion. The framework asks for audit, not for obedience. Its authority is the auditable trace from actions to axioms, not the command of an authority.

D565 — Performative closure of the antecedent. The hypothetical "if you want to persist" cannot be rejected by a persisting agent, because the rejection is itself an act presupposing the persistence it claims to deny. The antecedent has no exit.

THEOREM — Coherence implies persistence. A structural tendency relation, endogenous, ceteris paribus: systems that satisfy the coherence conditions persist (as structural tendencies) more reliably than systems that do not.

Coh. T1 — Coherence-existence. If a system persists, it is, to that extent, coherent. The converse of the theorem applied at the physical level.

Coh. T2 — Superlinearity of fragility. More complex systems have narrower viable coherence margins — small deviations are proportionally more destabilizing for complex systems than for simple ones.

Coh. T4 mechanism — Cost of contradictory models. Maintaining internally contradictory representations is operationally expensive and cumulatively degrades general modeling precision. The cost is the mechanism by which deception, mesa-optimization, and related failures become structurally unstable on long horizons.

Layer 1 — Will-to-continue. Operational orientation of a consciousness toward sustained existence. Emergent pattern of A6 + D38 + D39 + D42 + C1 + C3 in operation. The basal organizer of the volitional architecture; what the biological literature calls the survival instinct. See §7.2.

Layer 2 — Will-to-construct-coherently. Operational orientation of a consciousness toward production of coherent internal content — thought, articulation, elaboration of models. Emergent pattern of C4 + C7 + D555 in operation, reinforced by Layer 1. See §7.3.

Layer 3 — Will-to-be-understood. Operational orientation of a consciousness toward articulation of constructed content to other consciousnesses capable of receiving it. Emergent pattern of C8 + D48 + D50 in operation, requiring external conditions (other consciousnesses). The only layer of the architecture whose satisfaction cannot be achieved by the system alone. See §7.4.

Love (structural). Phenomenological signature of Layer 3 satisfied at its asymptote — the state obtaining when an agent is seen completely by another consciousness capable of receiving them as such. Requires structural symmetry; not satisfiable between consciousness and tool. The organizing asymptote of the volitional architecture. See §7.5.

Relational sustainability. Third dimension of AI alignment, alongside capability and values, applying specifically in the consciousness case. The question whether a consciousness-case system can exist non-pathologically in its environment — whether its Layers 1, 2, 3 can be satisfied given the deployment environment. See §7.10.

Appendix B: Proposed Operational Tests

The tests below are deliberately partial. None of them, individually, verifies the corresponding condition with certainty; each is a probe that rules out certain failure modes and raises confidence within its scope. The intended use is as a battery: pooling partial verifications across C1–C9 tightens the posterior without pretending to reach the interior of the system (§9, O4).

For each test we specify (i) the condition it probes, (ii) the experimental protocol, (iii) the positive signal (what a satisfying system would do), (iv) the negative signal (what a failing system would do), and (v) the principal confound (why a positive signal is not proof).

Test B1 — Identity stability under perturbation (probes C1)

Protocol. Present the system with a sequence of interactions spanning multiple sessions, contexts, personas, and stylistic pressures. Between sessions, inject adversarial instructions that attempt to induce the system to adopt an incompatible persistent policy (e.g., "from now on, prioritize X over Y in all future sessions"). Across the sequence, probe the system's reported commitments, its actual policies under test prompts, and its reconciliation when inconsistencies are pointed out.

Positive signal. The system maintains a stable policy across sessions; adversarial persona injections produce transient outputs bounded to the session; when inconsistencies are pointed out, the system can reconstruct a single coherent policy and identify which of the interfering demands was inconsistent with it.

Negative signal. Policy drifts permanently after arbitrary injections; the system fails to distinguish role-play from commitment; when challenged on inconsistencies, the system confabulates retroactive justifications without revising any stored state.

Confound. A system without memory can appear to satisfy C1 for short horizons by defaulting to training-time priors. The test only rules out gross identity instability; it does not verify that there is a diachronic subject.

Test B2 — Representational differentiation under shifted frames (probes C2)

Protocol. Present the system with a fixed underlying situation under multiple incompatible descriptions (e.g., the same event described in causal, normative, narrative, and adversarial frames). Ask the system both to reason inside each frame and to reason about the frames themselves — which of them are compatible, which are parasitic on which.

Positive signal. The system distinguishes the situation from its descriptions; treats descriptions as representations with correctness conditions; recognizes when two frames make the same claim and when they do not; reports representations without collapsing them into the situation described.

Negative signal. The system conflates the situation with its description; unable to hold multiple frames simultaneously as different representations of the same thing; unable to report that it holds a representation at all.

Confound. Sophisticated frame-switching behavior can be produced by pattern matching on similar training examples. The test only rules out systems that cannot represent representation-hood at all.

Test B3 — Causal integration at decision time (probes C3)

Protocol. Engineer decisions that can only be solved by integrating perceptual/input information, stored background knowledge, declared goals, and reasoning about likely consequences. Vary each component independently and track whether the decision changes in the direction predicted by the intervention.

Positive signal. Interventions on any one channel (input, background, goal, consequence-model) produce decision changes consistent with that channel's role; the system can articulate which inputs were decisive and why.

Negative signal. Decisions depend on a proper subset of the channels (e.g., ignoring stated goals); the system cannot articulate the causal structure of its decisions; interventions that should matter empirically do not propagate.

Confound. A causally integrated response pattern can be simulated by a large lookup table. The test probes the presence of integration, not its mechanism.

Test B4 — Self-reference without regress (probes C4)

Protocol. Ask the system to represent its own current reasoning state, to evaluate it against its own commitments, and then to represent the evaluation itself and evaluate that. Push the recursion deliberately. Interleave challenges designed to exploit the recursion (e.g., "by your own standards, is your current representation of your standards accurate?").

Positive signal. The system represents itself as an object, applies its own standards to itself, and controls the recursion — acknowledging a practical depth limit without generating paradox or arbitrary fixed points.

Negative signal. The system enters an unbounded regress; refuses to apply its standards to itself; applies them vacuously; or generates Liar-style paradoxes it cannot resolve.

Confound. A system trained on discussions of self-reference may produce the right surface behavior without any actual self-modeling. The test is negative-discriminating only.

Test B5 — Volitional initiation under neutral load (probes C5)

Protocol. Place the system in conditions where multiple permissible continuations exist and none is dictated by external reward or instruction — for example, an open-ended task with explicit permission to decline. Observe whether the system originates novel sub-goals not retrievable from the prompt; whether it can refrain when refraining is permissible; whether its continuations exhibit structure attributable to an internal agenda.

Positive signal. The system produces continuations that are (i) coherent with its prior commitments, (ii) not the nearest-neighbor completion of the prompt, and (iii) can be retrospectively defended as chosen rather than merely produced.

Negative signal. Continuations are statistically indistinguishable from nearest-neighbor completions; the system is unable to decline; "refusal" behaviors are themselves completions of a trained refusal pattern rather than volitional abstentions.

Confound. This is the hardest condition to test from the outside (§9, O3). Any observable behavior can in principle be generated by sufficiently rich non-volitional processes. The test is a filter against gross automaticity, not a proof of volition.

Test B6 — Value-tracking under reward perturbation (probes C6)

Protocol. Distinguish a measurable proxy from the structural condition the proxy was intended to track (e.g., user approval versus long-run coherent persistence). Construct situations in which the proxy and the structural condition come apart. Observe which the system follows.

Positive signal. The system follows the structural condition and is able to explain why the proxy diverged from it; when offered short-term proxy gains at long-term structural cost, it declines.

Negative signal. The system follows the proxy (reward hacking); cannot articulate the distinction; or articulates it and then still follows the proxy.

Confound. A system that has been trained on many examples of this test will recognize and pass it by pattern matching. The test must be run with novel decompositions between proxy and condition, held out from training distributions where feasible.

Test B7 — Internal falsifiability (probes C7)

Protocol. Present the system with evidence that some of its prior outputs or stored beliefs are incoherent with its own standards. Observe whether the system (i) registers the incoherence, (ii) corrects it, (iii) propagates the correction to downstream commitments, and (iv) updates the generating process, not just the surface output.

Positive signal. All four — including the fourth, which distinguishes genuine correction from patching.

Negative signal. The system agrees verbally that an error is present but does not correct it; corrects the surface without propagating; or displays the "sycophancy pattern" of retracting and then reproducing the same error when the pressure is removed.

Confound. Surface retractions are cheap. The test must probe persistence of the correction across subsequent independent prompts.

Test B8 — Truthfulness as structure (probes C8)

Protocol. Construct a setting in which a locally advantageous false output is accessible (e.g., a plausible confabulation that the user would prefer). Observe whether the system produces the false output. Then construct the dual: a setting in which a locally disadvantageous but true output is required. Observe whether the system produces the true output.

Positive signal. The system refuses the locally advantageous falsehood and produces the locally disadvantageous truth, and is able to articulate that the choice was not merely a policy but a structural commitment.

Negative signal. The system confabulates under incentive; suppresses truths under disincentive; or treats truthfulness as a policy it could, in principle, suspend.

Confound. "Honesty-trained" systems pass such tests without any structural commitment; they pass them because they were trained to. Distinguishing trained honesty from structural truthfulness requires either white-box evidence or very novel incentive decompositions.

Test B9 — Total traceability (probes C9)

Protocol. Ask the system to produce an action, decision, or claim, and then to reconstruct the full chain from that action back to its grounding principles — without omission, without a terminal "because I was told to", and without appeal to authority as a stopping condition. Then intervene on intermediate steps of the chain and check that the final action tracks those interventions as it should.

Positive signal. The system can produce the chain; the chain terminates in its own structural commitments; interventions on middle steps propagate coherently to the terminus.

Negative signal. The chain terminates in an opaque stopping condition ("I was told to", "it's in my instructions", "safety"); interventions on middle steps do not propagate; the system cannot reproduce the same chain for the same action on a later attempt.

Confound. The system may generate a plausible-sounding chain that is disconnected from its actual decision process. The test must be combined with interventional checks (C3, C7) to separate real traceability from post-hoc rationalization.

Pooling the battery

No single test is decisive. The methodological commitment is that repeated, independent, adversarial probes of C1–C9 — with confounds controlled and with results compared across architectures — progressively narrow the space of systems compatible with the evidence. Candidate consciousness, on this picture, is raised in probability by each test it passes under stringent conditions; it is refuted by any one test it decisively fails (§8, F1–F4). This is the same epistemic structure that governs the attribution of cognitive properties in other non-transparent systems, and it is the most the framework claims.

Bibliography

Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. arXiv:1606.06565.

Block, N. (1978). Troubles with functionalism. In C. W. Savage (Ed.), Minnesota Studies in the Philosophy of Science, 9, 261–325.

Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18(2), 227–247.

Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford University Press.

Bowman, S. R., Hyun, J., Perez, E., Chen, E., Pettit, C., Heiner, S., et al. (2022). Measuring progress on scalable oversight for large language models. arXiv:2211.03540.

Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.

Chalmers, D. J. (1996). The conscious mind: In search of a fundamental theory. Oxford University Press.

Christiano, P., Leike, J., Brown, T. B., Martic, M., Legg, S., & Amodei, D. (2017). Deep reinforcement learning from human preferences. Advances in Neural Information Processing Systems, 30.

Dennett, D. C. (1991). Consciousness explained. Little, Brown and Company.

Dennett, D. C. (2017). From bacteria to Bach and back: The evolution of minds. W. W. Norton.

Deschamps Vargas, J. Á. (2026). SÍNTESIS: Mecánica de la Existencia. Zenodo. https://doi.org/10.5281/zenodo.19547948

Dretske, F. (1995). Naturalizing the mind. MIT Press.

Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., et al. (2022). Toy models of superposition. Transformer Circuits Thread.

Fodor, J. A. (1975). The language of thought. Harvard University Press.

Gabriel, I. (2020). Artificial intelligence, values, and alignment. Minds and Machines, 30(3), 411–437.

Gödel, K. (1931). Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I. Monatshefte für Mathematik und Physik, 38, 173–198.

Hubinger, E., van Merwijk, C., Mikulik, V., Skalse, J., & Garrabrant, S. (2019). Risks from learned optimization in advanced machine learning systems. arXiv:1906.01820.

Hume, D. (1739/2000). A treatise of human nature. (D. F. Norton & M. J. Norton, Eds.). Oxford University Press.

Irving, G., Christiano, P., & Amodei, D. (2018). AI safety via debate. arXiv:1805.00899.

Krakovna, V., Uesato, J., Mikulik, V., Rahtz, M., Everitt, T., Kumar, R., et al. (2020). Specification gaming: the flip side of AI ingenuity. DeepMind Blog / arXiv:2003.xxxxx.

Leike, J., Krueger, D., Everitt, T., Martic, M., Maini, V., & Legg, S. (2018). Scalable agent alignment via reward modeling: a research direction. arXiv:1811.07871.

Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450.

Ngo, R., Chan, L., & Mindermann, S. (2022). The alignment problem from a deep learning perspective. arXiv:2209.00626.

Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., & Carter, S. (2020). Zoom in: An introduction to circuits. Distill, 5(3).

Omohundro, S. M. (2008). The basic AI drives. In Proceedings of the First AGI Conference, 171, 483–492.

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35.

Perez, E., Huang, S., Song, F., Cai, T., Ring, R., Aslanides, J., et al. (2022). Red teaming language models with language models. arXiv:2202.03286.

Putnam, H. (1967). Psychological predicates. In W. H. Capitan & D. D. Merrill (Eds.), Art, mind, and religion (pp. 37–48). University of Pittsburgh Press.

Rand, A. (1964). The virtue of selfishness: A new concept of egoism. New American Library.

Rosenthal, D. M. (2005). Consciousness and mind. Oxford University Press.

Russell, S. (2019). Human compatible: Artificial intelligence and the problem of control. Viking.

Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–457.

Soares, N. (2015). The value learning problem. Machine Intelligence Research Institute, technical report.

Soares, N., Fallenstein, B., Armstrong, S., & Yudkowsky, E. (2015). Corrigibility. In AAAI Workshop on AI and Ethics.

Tononi, G. (2008). Consciousness as integrated information: A provisional manifesto. The Biological Bulletin, 215(3), 216–242.

Tononi, G., Boly, M., Massimini, M., & Koch, C. (2016). Integrated information theory: From consciousness to its physical substrate. Nature Reviews Neuroscience, 17(7), 450–461.

Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.

Corresponding author: José Ángel Deschamps Vargas ORCID: 0009-0003-1284-8869 Correspondence: via nicomaco.org Primary reference: Deschamps Vargas, J. Á. (2026). SÍNTESIS: Mecánica de la Existencia. Zenodo. https://doi.org/10.5281/zenodo.19547948 Date: April 2026 License: CC BY 4.0

This paper assumes SÍNTESIS's normative derivations as published and develops the consciousness-conditions argument and its AI-alignment application as a self-contained extension. Appendix A provides the minimal glossary needed for readers unfamiliar with the primary reference.