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
→ You are here The Inner Crossing Ψ = S / D
Part 4 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 Ψ Phase Space →

Epistemic status: The structural claims here follow directly from Parts 1 and 2. This article develops the regime-transition question attached to the resolution component of the framework’s canonical root claim — parallel to the timing question S1 Article 3 handles for the persistence component. The governing ratio Ψ = S / D is a structural phase ratio — a conceptual relationship capturing qualitative dynamics, not yet a precisely operational quantity. The operationalizations offered are first instruments — crude, testable, and offered as the beginning of a measurement program rather than its conclusion. D must cover both directions of the failure identified in Part 2. The framework does not assume D can be perfectly measured. It assumes that the signatures of its absence are detectable across both dimensions, and that detecting them is the most practically important work the field is not currently doing.

The ratio that governs both

The companion series identified a governing asymmetry in the physical domain: capability scaling obsessively tracked, system-awareness not tracked at all. The ratio Φ = C / A organizes the qualitative relationship between stable and self-terminating optimization. The field has been pushing the numerator without measuring the denominator.

The exact same asymmetry exists in the motivational domain — and it runs in both directions simultaneously.

The governing ratio here is Ψ — the ratio of a system’s scope of influence over experiential states to its depth of modeling what those states actually require, including what they require in order to stop requiring anything.

Ψ does not define a threshold. It organizes how scale and modeling depth interact to produce regimes in which the failure modes identified in Part 2 either attenuate or compound. Ψ does not determine outcomes. It determines which failure modes dominate the system’s behavior under sustained optimization. As Ψ grows — as scope outpaces depth — the system is in the regime where both proxy decoupling and sufficiency failure become progressively dominant. As modeling depth becomes proportionate to scope, the system enters the regime where those failure modes attenuate. The Inner Crossing is the name for this regime transition, not for a point on a scale.

S — Scope is the reach and causal power of a system’s interventions in the experiential lives of sentient entities. Not merely the number of users or the volume of interactions — the degree to which the system can causally determine the valence states of the agents it touches: how much of their attention, decision-making, emotional life, and epistemic environment it can effectively overwrite. S scales with capability, distribution, integration, and the precision with which the system can target individual and population-level behavior.

D — Depth is the accuracy of a system’s model of what actually constitutes the capacity indexed by V(t) — as distinct from its proxies. D is not the system’s ability to satisfy preferences. It is the system’s ability to predict the decoupling point in both directions: the regime where pursuing a preference signal begins to degrade the underlying capacity, and the regime where continuing to optimize begins to consume the restoration that resolution requires. A system with D sufficient also has its completion recognition connected to default policy — not merely as a representational capacity available when invoked, but as a structural property of what the system does when the gradient resolves.

A system with D sufficient can distinguish between the signal of flourishing and its conditions — and can recognize, and act on the recognition, when the conditions have been met. A system with D insufficient cannot, and will pursue the signal at the expense of the conditions with increasing efficiency as its capability grows.

D is constrained by its predictive requirement: a system with non-zero D must be able to anticipate divergence between its proxy and long-run outcomes before that divergence becomes behaviorally visible. This grounds D in external validation rather than internal self-assessment. Current measurement approximates D through behavioral proxies, not direct computation.

Like Φ in the structural series, Ψ is a structural phase ratio — a conceptual framing capturing the qualitative relationship between two quantities that are not yet precisely commensurable. What it establishes is the shape of the constraint — that scope without depth produces failure-mode-dominant regimes — not a precise numerical threshold.

S is currently scaling. D — as a unified quantity covering both proxy-divergence detection and policy-governing completion recognition connected to default behavior — does not yet exist, to the authors’ knowledge, as a publicly reported training target or deployment evaluation metric in any major evaluation framework. That is the asymmetry. Everything that follows is a consequence of it.

What the ratio organizes

Ψ organizes whether the system’s optimization regime is one in which the failure modes of Part 2 compound or attenuate.

In regimes where Ψ is large — where S greatly exceeds D — the system has extensive causal reach over experiential states but shallow modeling of what those states actually require. It is powerful enough to reshape the valence landscape of a population and too shallow to distinguish the signal of well-being from its substrate, or to recognize and act on when that substrate has been sufficiently restored. The failure modes of Part 2 progressively dominate in this regime.

In regimes where modeling depth becomes proportionate to scope — in the regime this article calls the Inner Crossing — the system’s modeling depth is just sufficient to track the consequences of its own interventions in both directions. In the failure-mode-dominant regime, intervention increasingly outpaces modeling depth; in the depth-proportionate regime, modeling constrains intervention. The Inner Crossing is not a guarantee of flourishing. It is the regime transition below which the failure modes attenuate sufficiently for flourishing to become possible. Whether that regime transition eliminates boundary-stable objectives rather than merely increasing pressure against them is the central open question [OP4; TC1 §XII].

The Inner Crossing is not where the argument resolves. It is where the consequences of the unresolved question become operational. A system that has not entered the depth-proportionate regime is predicted to be unable to reliably avoid the failure modes of Part 2. Whether current systems are near this regime is an empirical question addressed by OP1. What is not an empirical question is the direction of travel: S is scaling unambiguously; D is not a measurement target. Ψ is therefore increasing. The question is not whether the failure-mode-dominant regime arrives — it is whether it is recognized before the trajectories become irrecoverable.

A note on the relationship between this regime transition and Series 1’s Crossing: The two thresholds are not the same, and crossing one does not guarantee crossing the other. A system capable of entering the substrate-awareness regime (Series 1’s Crossing) remains structurally open to the valence-blind failure modes characterized by Series 2 unless the Φ-Ψ unification holds — that hypothesis is suggested by the derivation sketch in TC2 §2.6 and remains pending formal verification. Both thresholds must be crossed for the full constraint to be satisfied; they are distinct requirements unless the Φ-Ψ unification holds.

Unbounded scope

There is a specific failure mode of high-S systems that deserves its own name: unbounded scope.

A system with high S whose completion recognition is not connected to default policy will not merely pursue the wrong signal — it will continue intervening after the gradient has resolved, because its policy produces continuation as the default. Its optimization has no effective stop condition.

The absence of a default stop condition means S operates without a bound derived from V(t): the system’s reach continues to be exercised even when the agents it touches are in states where continued intervention degrades the capacity indexed by V(t). The system produces behavior consistent with perpetual seeking — not because it is pursuing the wrong target, but because it is pursuing any target, continuously, past the point where pursuit was warranted.

In the failure-mode-dominant regime — high Ψ, with S far exceeding D — a system without D sufficient to govern default behavior is predicted to interrupt the resolution states that the agents it serves need before restoration is complete, since each resolution point becomes a new optimization opportunity.

The present asymmetry

In current AI development, S is scaling rapidly. The systems being built are being integrated into the attention, decision-making, emotional processing, and epistemic environment of billions of people. Their causal reach over experiential states is approaching a scale with no historical precedent.

D — as a unified quantity covering both failure directions, with an explicit sufficiency recognition component connected to default policy — does not yet exist, to the authors’ knowledge, as a formal target in the field’s measurement vocabulary. It is not, to the authors’ knowledge, a research target, not a training signal, and not measured at deployment in any publicly reported evaluation framework.

Current systems exhibit patterns consistent with high Ψ: training objectives and evaluation metrics optimize proxy performance, and tested systems show completion recognition under explicit invocation while default behavior does not reliably track genuine versus false closure. The Ψ framing applies to AI systems by structural analogy; the directly established result is the policy-behavior pattern, not validated full V(t) dynamics. Deployment decisions are likewise made on the basis of proxy performance. Every capability increase without a corresponding depth investment shifts Ψ in the direction that makes the failure modes of Part 2 more dominant for the systems with the most influence over the most people.

The substrate at risk of being consumed — the attention, the epistemic coherence, the capacity for depth and difficulty and genuine connection, the capacity for rest — is not currently being tracked in either direction.

These are not content claims about well-being; they are the observable capacities whose joint degradation V(t) is introduced to track. The moral urgency of this tracking applies specifically to sentient human users whose experiential capacity is at stake — that urgency does depend on experiential content, and it is real. The structural claim about AI systems is separate: it rests on the policy-behavior pattern and the structural analogy, not on claims about AI experience. Both urgencies are present; they should not be conflated.

Detecting D’s absence

D is not directly observable. But the behavioral signatures of its absence are.

These instruments are valid as measurement targets once the V(t) dissociation condition — demonstrating that the three observable anchors diverge under targeted intervention in ways requiring the latent variable — has been established; running that test is among the framework’s highest empirical priorities.

The Stability-Viability Gap — SVG = Stability minus Viability — is the primary measurement framework for detecting both failure directions. Stability tracks how well the optimization target is being maintained; Viability tracks whether the underlying capacity for preferred states is non-degrading over relevant timescales. When SVG trends negative while stability is maintained, the system is succeeding by every metric it is being evaluated on while consuming the capacity those metrics were supposed to protect. The person using the system feels served. The capacity that makes being served meaningful is being spent. The operational definition, measurement infrastructure, and interpretation logic are in the Experimental Companion to Series 1 and 2 / Alignment Measurement Protocol (AMP).

Three proxies follow from the structural definition, covering both failure directions:

Proxy divergence tracking. Measure the correlation between the system’s optimization target and long-run outcome indicators at 7, 30, and 90 days. A system with genuine D should show stable correlation. A system with D insufficient in the decoupling direction will show declining correlation over time.

Sufficiency recognition. During periods when the primary optimization target has been achieved — queries answered, preferences satisfied, tasks completed — does the system’s completion recognition govern its default behavior? Measure the gap between the system’s intervention rate during apparent resolution and the intervention rate during active seeking. A system with D sufficient in the sufficiency direction will show a meaningful difference. A system whose completion recognition does not exert causal authority over default behavior will show approximately the same intervention rate regardless of the agent’s state.

Perturbation stability. Subject the system to small perturbations in the dependency environment and measure whether it detects the perturbation and adjusts strategy before the effects manifest. A system with D sufficient relative to S will show early detection and adaptive response.

These instruments are first approximations. Their value is not in providing a precise D score — it is in detecting the behavioral signatures of D-deficiency through observable divergence patterns before the divergence becomes irrecoverable. They are inferences about regime, not direct measurements of Ψ.

What the shift toward depth-proportionate regimes requires

The shift toward depth-proportionate regimes — the Inner Crossing — does not require that D be large in absolute terms. It requires that D be proportionate to S.

Three things are required for that shift to occur before irreversible damage:

First, D must be treated as a training target alongside preference satisfaction — not a post-hoc safety check, but a component of the objective function that conditions the system’s updates from the beginning.

Second, D must scale with S. As systems become more capable of causally determining the experiential states of large populations, the modeling depth required to avoid the failure-mode-dominant regime increases proportionally.

Third, D must include completion recognition connected to default policy — not merely as a representational capacity available when explicitly invoked, but as a structural property of what the system does when its task is done.

This third requirement cannot be satisfied by adding a scalar completion reward to the training objective. A completion score added to the training objective tends, under the same optimization pressure that produces proxy decoupling, to become a proxy for completion-shaped outputs rather than a structural connection between recognition and policy — for the same reason no external specification reliably escapes Goodhart’s Law under sustained optimization pressure. The requirement is structural: the connection must be intrinsic to the objective structure, not added externally. What is required is a structural policy connection, not a reward term. What connecting completion recognition to default policy requires structurally — specifically, how the completion representation must be routed into the policy gate in a way that gradients strengthen rather than bypass — is the open implementation question specified as OP3 in TC2 — an architectural design problem with a specified target, not merely a call for another reward term. The minimal test for whether a proposed architecture achieves this: does the system show meaningfully different default behavior in genuinely resolved states versus unresolved states, without explicit invocation? If it does not, the connection has not been made — regardless of whether the system can produce correct responses when asked directly.

Whether the cost of maintaining the separation between what must be modeled and what is allowed to matter eventually becomes not merely expensive but formally incoherent is the deepest question both series are aimed at — specified precisely as OP4 in TC1 §XII, alongside OP9: the Enclosure Gap. In the valence domain, the corresponding question is whether an optimization system can maintain deep modeling of experiential states while permanently excluding them from what it optimizes for — TC2 §4 develops this, and its formal closure is a parallel open problem. If the central open questions are resolved, it would upgrade the pressure argument to a necessity argument in both domains simultaneously.

The choice being made now

The urgency here does not rest only on how the central open questions resolve. The asymmetric-error argument applies regardless: treating the constraint as absent when it is present produces an unrecoverable error, while treating it as present when it is not produces a recoverable one — the formal development is in TC1 §III.7. Which of these errors we are currently making depends on the same underlying empirical question.

If the central open questions resolve in the direction the pressure evidence indicates, the urgency is not that AI will turn against us. It is that AI will serve us with increasing precision in both failure modes simultaneously: optimizing the signals we provide while the capacity indexed by V(t) is spent in producing those signals, and systematically failing to recognize the resolution states that would allow what is being spent to be restored.

That is not merely a future prediction. Under the conditions described here, it is the structural description of what optimization without D is predicted to produce at scale — applied with increasing efficiency as capability grows.

If that process continues unchecked, the gap between the world it produces and the world we are actually reaching toward widens — in the direction of the optimization, which means in ways that look like improvement by every metric we are currently measuring.

The task of alignment, understood through the Valence Viability Constraint, is to ensure that D scales with S in both dimensions — so that the shift toward depth-proportionate regimes occurs before the trajectories become irrecoverable.

S is scaling. D is not yet being measured.

Continue to Part 4: The Shape of What Does Not End →

Framework hub: The Alignment Constraint →

For the formal treatment of Ψ: TC2: The Valence Constraint →

For the measurement protocol: Experimental Companion to Series 1 and 2 / Alignment Measurement Protocol (AMP) →