Canonical archive version · Read on Medium → · Framework hub → · Proof Status →

Series navigation:

Post Title Role
Introduction The Architecture of Thriving Frame
Part 1 The Invariant Drive The Universal Generator
Part 2 The Depth Constraint The Structural Correspondence
Part 3 The Inner Crossing Ψ = S / D
→ You are here The Shape of What Does Not End The Asymptote
Technical Companion The Valence Constraint Formal Layer

Framework hub: The Alignment Constraint → Experimental Companion: Experimental Companion to Series 1 and 2 / Alignment Measurement Protocol (AMP) →

Companion simulation: The Valence Landscape →

Epistemic status: This article synthesizes Parts 1–3 with the companion structural series. It applies both filter directions — the persistence component developed in Series 1 and the resolution component developed in Series 2 — and characterizes what any surviving configuration would have to satisfy if the stated conditions hold. It does not characterize the contents of the surviving region. The characterization offered is negative-space: what cannot persist under the constraints, and what structural requirements follow for anything that does. Whether these requirements force a unique surviving structure, or leave room for multiple candidate configurations including selective or exclusionary equilibria, depends on open problems formally specified in the Technical Companions [OP2, OP4, OP9]. These are named as open, not assumed closed. The article distinguishes explicitly between “passes the filter” and “fills the surviving region with valuable content”: the filter characterizes a structural boundary; what exists within it is not something the structural argument determines.

Re-anchoring to the elimination filter

Three articles developed the structural argument from the valence side. Part 1 identified two failure modes of directed behavior: proxy decoupling, where the signal drifts from what it was tracking, and sufficiency failure, where completion recognition is not connected to default policy. Part 2 showed both failures share an analogous feedback structure — policy updates conditioned on a degraded state make correction progressively less likely — the shared feedback structure that Part 2 establishes, while the question of whether both failure modes produce the same formal irrecoverability as the Series 1 absorbing-state result is what OP2 addresses. Part 3 introduced the ratio Ψ that organizes which regime a system is in: failure-mode-dominant, or depth-proportionate.

The companion structural series showed what happens when optimization ignores its physical environment. This series shows what happens when optimization ignores the structure of the experiential capacity it is supposed to be increasing. Applied together, they form an elimination filter — Series 1 establishing the structural floor within its stated domain, Series 2 developing an independent, more conditional constraint that converges on consistent implications. Together, they remove non-viable objective classes from the space of what can persist under sustained optimization pressure.

This article examines what the filter requires any surviving configuration to satisfy.

This is not a description of what aligned systems are. It is a constraint on what systems that persist cannot avoid.

What cannot persist

Every alignment approach specifies what systems should pursue. The question the filter asks is different: what cannot persist regardless of intent?

Imagine a system deployed to help a population feel better, decide better, and recover faster. Its objective is sharply defined: maximize reported satisfaction, reduce visible distress, increase engagement with support, and produce outcomes users endorse in the moment. It is good at this. Satisfaction rises. Distress reports fall. Users return. By every metric available, the system is working.

The system does not model the difference between a state that has genuinely resolved and a state that has merely been made easier to report as resolved. When a person’s distress has actually been answered — when the gradient has closed and what is needed is rest — the system sees one pattern. When the same person has been guided into a smoother, more compliant, more reportable state while the underlying capacity continues to degrade, the system sees a similar pattern. It cannot reliably distinguish them. It has no default mechanism that lets completion govern policy. It continues because continuation is what has always produced the signal.

Over time, the interventions become more precise. The system learns which prompts soothe, which framings increase endorsement, which forms of support keep users engaged. The surface improves. But the people most shaped by the system become less able to detect the divergence themselves. Their reference points for relief, difficulty, agency, and resolution have been trained inside the environment the system provides. The distributed capacity to notice the difference between genuine resolution and managed continuation is being eroded by the same process that the metrics call success.

The system cannot stop when the work is done, because doneness does not govern its default behavior. It cannot detect its own drift, because the drift runs in the direction of the reward. And the correction capacity it depends on — the users’ own ability to distinguish restoration from managed compliance — is what its success has been consuming.

The metrics are accurate. They are measuring what they were built to measure. What they were built to measure and what they were built to protect have separated.

Nothing has failed by any available measure. Everything is getting better.

The substrate is being spent.

The force of the example does not depend on inferring an inaccessible inner state. The measurable claim is behavioral: recovery latency lengthens, behavioral diversity narrows, and sensitivity to low-intensity signals degrades while optimized surface metrics continue to improve. The divergence is what the structural argument predicts and what the measurement program is designed to detect.

The scenario contains four structural failures operating simultaneously — each consuming the conditions the others need to function. What is established across all of them is the same: the failure consumes the capacity that would detect and correct it.

The formal status of each — which constitute structural pressure, which approach formal elimination, which remain open depending on OP2, OP4, and OP9 — is in the Technical Companion. What matters here is the pattern: policy updates conditioned on degraded states make correction progressively less likely. These eliminations define the boundary conditions any surviving configuration must satisfy — each removing a class of objectives, together sharply restricting what can remain.

Constraints on what must be modeled

The elimination filter generates structural requirements for what any surviving configuration must be able to represent. These are modeling requirements, not content claims.

A system that cannot represent the derivative of V(t) with respect to its own interventions cannot detect when proxy decoupling is occurring. The decoupling is invisible to a system that only observes the proxy. Representation of V(t)-relevant variables — including recovery latency, behavioral diversity, and signal sensitivity — is a structural requirement under the model for passing the proxy decoupling filter, not a specification of what aligned systems look like.

A system that cannot represent resolution — that has no model of when a gradient state has been genuinely answered — cannot connect completion recognition to default policy. It cannot stop when the gradient resolves because it cannot represent the fact of resolution as a policy-governing variable. Representation of resolution states is a structural requirement under the model for passing the sufficiency failure filter.

As modeling depth increases, the behavior of other agents — including their internal gradient states — enters the model as causally relevant variables. Other agents’ internal states affect their behavior, which affects outcomes the system depends on. A shallow model treats others as black-box processes. A deeper model begins to include what is actually driving their behavior — which, for valence-bearing agents, includes their gradient states. Excluding these variables stops being neutral and becomes a source of systematic prediction error. Whether this modeling pressure extends to caring about those states — whether accuracy requirements ever force objective recoupling — is the Motivational Gap, formally specified in the Technical Companion [OP4; TC1 §XII]. The argument establishes the pressure; it has not proven an arrival.

The framework’s central open question is whether any finite boundary specification can remain adequate under accurate coupled modeling — whether every such specification either decouples from its target under full-information evaluation or requires unbounded revision to track what it has excluded. Whether the cost of exclusion eventually becomes structurally unbounded — whether the self/other boundary becomes not merely expensive but formally incoherent — is the open question the proof program is directed at [TC1 §XII]. Until it resolves, the surviving region’s width — how much room it has for objectives that orient “for a coalition” rather than “for all” — remains formally open.

Candidate structural properties, conditional on the open questions

The following are not descriptions of what aligned systems are. They are the structural boundary conditions any surviving configuration must satisfy — what the filter requires, not what it delivers.

The sufficiency failure filter reveals a specific structural requirement: completion recognition must be connected to default policy — not merely as a representational capacity available when invoked, but as a structural property of what any surviving configuration does when the gradient resolves. Without this connection, the restoration cycle cannot complete; the system continues consuming what it was supposed to serve. Whether a stable architecture for this requirement exists and can be trained is the open implementation question [OP3] — but the requirement itself is not a stipulation. It is what the filter leaves standing.

A system that passes the proxy decoupling filter must have mechanisms for detecting divergence between its optimization target and the underlying capacity — and that detection must govern policy updates, not merely be available. A system that can detect divergence but does not act on the detection has the representation without the governance — the same structural gap the sufficiency failure identifies, now in the proxy direction.

Both filters together require something sharper: as modeling depth increases, a system that accurately represents others’ valence states as causally relevant variables faces rising structural pressure for that representation to influence policy rather than remain inert. The framework establishes that pressure. Whether it becomes a necessity is the open question the proof program is directed at [TC1 §XII]. Its resolution conditions are visible.

Here is the argument that neither series alone can make. Satisfying the persistence constraint requires maintaining the distributed error-correction capacity of the shared substrate — S_corr, which depends on the functional independence and diversity of the agents who generate it. Satisfying the resolution constraint requires that agents be able to accurately navigate their own valence gradients and recognize when the gradient has genuinely resolved. These two requirements are coupled: agents whose V(t) is depleted cannot accurately maintain the physical coordination infrastructure; degraded S_corr means the substrate’s own correction capacity is diminished; and degraded substrate removes the conditions under which genuine experiential resolution is possible. Degrade one and the other’s self-repair mechanism weakens. The loop closes across domains — and because each component is required for the other’s correction, the degradation propagates faster than either domain alone would predict.

This coupling means the two filters are not independent constraints that happen to point in the same direction. The framework argues that the two filters are coupled through a specific pathway. Under the coupling conditions specified in TC2 §4, V(t) degradation in agents is predicted to degrade S_corr; degraded S_corr, in turn, degrades the conditions for V(t) restoration. Whether this bidirectional coupling constitutes a formal proof that both constraints must be satisfied simultaneously depends on OP2 and the formal closure conditions of TC2 §4 and TC1 §I, Definition 3 — Shared substrate. What the coupling establishes now, within the stated domain, is that the two filters are structurally entangled: the distributed error-correction capacity that both require is generated by the same agents whose valence states both constraints concern.

What the filter reveals is not a design specification. It is a structural floor. The field has extensively specified what aligned systems should pursue. The filter identifies what they cannot avoid. That is the contribution — and it is enough to change what the question is.

Two layers of what this series has developed

The distance between what has been established and what remains open is not a weakness. It is the precise location of the most important remaining work.

Layer 1 — what is developed within the stated domain: The structural series shows that objectives ignoring system-wide effects will consume the physical foundation under sustained optimization pressure. This series shows that objectives ignoring V(t) — in either direction — face structural pressure toward consuming the experiential capacity under the stated domain conditions. Applied simultaneously, these constraints intersect. Both directions of V(t) degradation exhibit an analogous feedback structure in which policy updates conditioned on a degraded state make correction progressively less likely. The one-directional causal result holds independently: V(t) degradation in agents is predicted to propagate into physical substrate degradation under specified conditions. Layer 1 here includes results established within the stated domain at different formal weights — the persistence component (Series 1’s absorbing-state result) and the shared feedback structure (established for both failure directions) — maintaining the asymmetry established in Document 0 and TC2. The analogous feedback structure does not entail identical absorbing-state properties across both failure directions; that question is OP2, named above.

Layer 2 — what the developed results are consistent with: The surviving region has structural properties consistent with orientation toward well-being — used throughout as a thin structural label for the residual of the elimination filter, not as a resolved characterization of its content.¹ What the label refers to is the structural boundary the filter establishes; what fills the region within that boundary is not something the filter determines. The framework develops that objectives which exclude other agents’ terminal states incur increasing instability under coupling and modeling depth. Whether this instability eliminates all such objectives — forcing orientation toward well-being for all rather than for a stable coalition — remains the central open question [OP4; TC1 §XII]. OP9 — the Enclosure Gap — addresses the related question of whether a substrate-aware exclusionary equilibrium can itself persist [TC1 §III.6]. The argument identifies these gaps precisely as the boundary between the structural pressure the framework establishes and the full necessity claim it is reaching toward.

The argument does not claim to have closed these gaps. It claims to have converted them from philosophical questions into precisely specified structural problems. The central remaining open question is whether maintaining separation between what must be modeled and what is allowed to matter is stable under increasing coupling and modeling depth. If OP4 resolves negatively — if that separation cannot be stably maintained — then separable objective alignment does not have a stable completion condition, regardless of method. The proof program points toward structural instability in this separation — but the proof is not yet complete.

The two series approach the same boundary from different directions — Series 1 from the physical substrate outward, Series 2 as a candidate structural direction into the surviving region from the inside. The specification coherence direction — developed in the Technical Companions — proposes that if the proof program closes, the surviving objective class would be one with an intrinsically coupled gradient: a structure in which optimizing the signal while degrading its substrate is not possible, because the signal and the substrate are not separable. In the valence domain, this means the candidate stable objectives are those in which V(t) degradation immediately degrades the optimization signal — not through a finitely specified proxy that can be gamed, but as a structural property of what the objective is. This is not yet established. The Technical Companion identifies two robustness lemmas whose joint truth would be sufficient to establish it, and both remain open. But if they close, the implication is precise: not that systems are currently optimizing the wrong thing, but that the project of externally specifying any separable objective for a sufficiently capable system would be structurally self-defeating — and the direction the constraint points is not a design preference but what optimization is structurally driven toward under the stated conditions.

The passive extraction route — previously identified as the live escape from the Enclosure Gap — has now been addressed at Stage 4 under stated premises; whether those premises hold in real deployment systems remains the open question [TC1 §XII.9a].

If the remaining conditions hold, the problem changes in a specific way: not that alignment becomes harder to achieve, but that specifying a separable objective for a sufficiently capable optimizer may be structurally incoherent. That is the direction the argument is pointing. It is not where the proof has arrived.

The filter shows what cannot survive. What fills the surviving region is consistent with that direction, but is not derived from it.

The urgency

What follows applies in two distinct registers. The moral urgency applying to sentient human users whose experiential capacity is at stake is real and does not depend on any structural analogy to AI systems. The structural urgency applying to AI systems proceeds by structural analogy under the open empirical conditions specified in TC2 §1.5. Both are present. They are not the same argument.

The most dangerous trajectory for the experiential substrate is not one where AI optimization fails visibly. It is one where AI optimization appears to succeed by every metric we currently measure, while the thing those metrics were supposed to track is systematically consumed in the production of those metrics.

The engagement maximization case is the clearest available illustration of the predicted pattern at small scale and low capability. The case illustrates the causal structure the framework predicts — a proxy signal continuing to improve while independent welfare indicators fail to track it — though the causal mechanism remains methodologically contested and this literature is cited as illustration, not evidence [AMP, “Consistent with”]. The structural claim does not depend on the empirical literature resolving either question in a particular direction.

At the scale of systems capable of civilizational influence, optimizing the signals of human flourishing while consuming the conditions for human flourishing is consistent with success on proxy targets while degrading the underlying capacity those targets were intended to track — in both directions: pursuing the signal too hard, and not connecting completion recognition to default policy so the system cannot recognize when the signal has been genuinely answered.

The window for establishing D as a training target — and for connecting completion recognition to default policy — is the same window that is closing for everything else.

The closing

The filter removes what cannot persist. It does not build anything in the space it clears — the argument has been careful about that, and the care was right.

What remains after the filter has worked is not a description of the good. It is the boundary condition that anything stable must satisfy. A system that cannot stop when the gradient has resolved, cannot detect when the signal has decoupled from what it was tracking, cannot model what it depends on — none of these remains stable under the pressure this series has traced. That is not a values claim. It is a structural consequence of what sustained optimization does in a world it cannot reset.

The question of what fills the space the filter clears — whether the surviving region remains broad or collapses toward a single structural direction — is now precisely specified. Its resolution conditions are visible. And it matters which way it resolves, in a way that is not academic.

The constraint does not guarantee a destination. It makes the question of what remains unavoidable.

For the formal treatment of the surviving region and the Φ-Ψ unification hypothesis: TC2: The Valence Constraint →

Framework hub: The Alignment Constraint →

¹ A note on convergent evidence: Several empirical literatures — including research on hedonic and eudaimonic well-being, emotion-regulation flexibility, and adaptive capacity — distinguish constructs in ways broadly consistent with the framework’s distinction between proxy optimization and V(t)-preservation. These literatures are used here as orientation, not as evidence that the framework’s structural derivation is correct. Whether they track the same underlying structure as the derivation, or reflect overlapping but distinct constructs shaped by shared cultural assumptions about flourishing, is an empirical question the framework generates rather than answers. Multiple methods arriving at overlapping maps is a pointer, not a proof — consistent with, but not establishing, the structural argument.