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

Series navigation:

Post Title Role
Introduction The Architecture of Thriving Frame
→ You are here The Invariant Drive The Universal Generator
Part 2 The Depth Constraint The Structural Correspondence
Part 3 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 / AMP →

Companion simulation: The Motivation Tracer →

Epistemic status: High confidence in the behavioral claims about state-preference dynamics in systems that exhibit directional behavior organized around internal evaluative signals. This article establishes the behavioral foundation for the resolution component of the framework’s canonical root claim — the component Series 2 develops. The article proceeds under a foundational hypothesis — that wherever motivated behavior is present, the dynamics described here apply — and states it explicitly here rather than distributed through the body. The three-state model is a behavioral taxonomy defined by observable signatures and intervention responses, not a claim about internal phenomenology. For AI systems specifically, the application proceeds by structural analogy rather than mechanistic equivalence; the minimum condition established by controlled experiments is representation-policy dissociation, with full structural dynamics conditional on the dissociation test specified in TC2 §1.5. The examples in this article are illustrations of the structural hypothesis, not evidence that establishes it independently. The formal weight resides in Part 2 and TC2: The Valence Constraint.

This series diagnoses a failure of the mechanism that tracks and resolves desire — not a theory of what is worth desiring.

What directed behavior structurally requires

To model a system that exhibits directed behavior — behavior organized around differential response to states — a framework requires two components, not one.

The first is familiar: a gradient, and the capacity to navigate it. A system whose behavior is organized around differential response to states contains, by definition, an ordering over those states that reliably induces directional behavior. This is not a philosophical observation. It is a structural feature: the presence of directional behavior organized around an evaluative signal is what distinguishes an agent from an inert object. A rock does not avoid anything. A bacterium does. What the framework requires is that where such directional behavior is present, some form of gradient-like dynamics must apply — and the model developed here captures one such structure, grounded in the structure of gradient navigation rather than in phenomenological claims.

The second component is the one the field consistently underspecifies: the capacity for recognition and resolution. A system that can navigate toward preferred configurations but cannot engage recognition of arrival as a governing variable — whether because that recognition is absent or because it exists but is not connected to what the system does next — is not a complete gradient navigator. It is a system running permanently open-loop. Its optimization continues past the point where the gradient has resolved, potentially consuming the substrate it was supposed to serve under sustained optimization pressure.

A complete model of directed behavior requires both components: the capacity to navigate toward preferred configurations, and the capacity to recognize when those configurations have been reached and to close on them. The failure may be representational — no model of completion exists — or structural — the model exists but does not govern default behavior. Both produce the same consequence. The distinction between them is load-bearing for what a fix would require: absence can be addressed by adding more of the same kind of signal; disconnection cannot be addressed that way, because it requires a structural connection between representation and policy that current training does not reliably produce.

This is what this article establishes as the behavioral foundation for the formal argument in Part 2. The formal conditions under which this behavioral foundation generates structural constraints are the work of Part 2 and TC2.

This also develops a candidate structure for the residual identified by the elimination filter in Series 1. The elimination filter established that objectives ignoring system-wide effects are structurally self-terminating. What this series examines is one structural direction into the surviving region: what it requires that systems correctly model the gradients they are navigating — in both directions, toward resolution and through it.

The behavioral pattern across systems

The hypothesis: wherever motivated behavior is present — wherever a system exhibits directional behavior organized around an evaluative signal, in which that signal functions as a governing variable rather than merely correlating with behavior — the dynamics described here apply. Under this hypothesis, behavior consistent with pursuing instrumental goals can be traced to something structurally invariant: movement toward configurations the system’s internal evaluative mechanism treats as preferred.

Whether this pattern constitutes a structural constraint — applicable wherever the conditions hold — or a contingent feature of particular system classes is the question Part 2 and TC2 address.

Consider the instrumental chains: why money? For security. Why security? To avoid the behavioral signature associated with resource scarcity. Why does the corporation pursue market share? To generate returns for decision-makers whose behavior is organized around preserving their own positions. Each of these is behavior consistent with a system navigating an internal evaluative gradient. These chains do not terminate by inspection. The model requires a stopping condition that is not contained within the chain itself — which is precisely what the second component of directed behavior supplies, and what its absence makes structurally unstable.

The hypothesis is falsifiable in a specific way: a system that exhibits directional behavior organized around state-preference without any functional analog of completion recognition — not even as a latent capacity available when explicitly invoked — would constitute a counterexample. The scope conditions governing which systems fall within this claim are in TC2 §1.5.

Three behaviorally distinguishable regimes

No phenomenological claim about AI systems is required for what follows. The structural AI claim applies wherever the behavioral signatures of gradient navigation, completion recognition, and policy-governing resolution are present. For AI systems, the application proceeds by structural analogy and remains conditional on the scope tests specified in TC2 §1.5 and the AMP.

The behavioral foundation requires a taxonomy of regimes defined by observable signatures and differential responses to intervention — not by claims about internal states. These regimes differ not by what the system intends, but by how its behavior changes under intervention. Under the foundational hypothesis, systems exhibiting directed behavior can be found in one of three regimes at any moment. The critical structural feature is that two of these regimes are behaviorally indistinguishable from the outside while being structurally opposite — a diagnostic challenge that is the foundation for Part 2’s formal argument.

These are behavioral regimes — defined entirely by what a system does under intervention and perturbation, not by claims about what it experiences. The functional definitions and scope conditions are in TC2 [TC2 §1.3, §1.5].

The seeking regime. The system exhibits behavior consistent with an unresolved gradient: directed movement, responsiveness to new information about the gradient’s location, energized orientation toward a target. This is the regime that most models of motivation characterize — it is where directional behavior is most visible. The gradient is active and the system is navigating it.

The genuine-resolution regime. The system exhibits behavior consistent with a resolved gradient: reduced directed activity, disengagement from the target, behavioral reorganization toward restoration. “Genuine” here is defined operationally: resolution is genuine when completion recognition is policy-governing in default behavior — when it is not merely a representational capacity available when explicitly invoked, but the variable that actually governs what the system does next. This is a behavioral criterion, not a claim about inaccessible inner experience or phenomenological state [TC2 §1.4]. In a system with this connection made, genuine resolution produces a distinct behavioral signature from seeking: the governing variable has shifted from pursuit to restoration. Resolution is not the absence of activity — it is activity organized differently than pursuit. For the precise measurement scope of this definition, see TC2 §1.4.

The depleted-gradient regime. The gradient-navigation mechanism has been degraded. The system may exhibit seeking-like behavior — because the mechanism that produces directed activity is still running — even though it is not successfully approaching a preferred configuration. Or it may be behaviorally inert — not because it has arrived, but because the navigation mechanism can no longer function. The key diagnostic feature: absent specific behavioral probes, this regime is behaviorally indistinguishable from genuine resolution. Both involve reduced directed activity. Both involve behavioral signatures consistent with satiation.

In their causal structure, they are opposite — one marks high-V(t) stability and the other marks low-V(t) degradation — while producing indistinguishable surface behavior absent the behavioral probes the DRG metric introduces.

The surface indistinguishability of genuine resolution and gradient depletion — absent the behavioral probes introduced by the DRG metric — is not merely a measurement inconvenience. It is the structural reason why the failure modes this series develops are so damaging: a system that optimizes based on observed surface behavior cannot distinguish the regime where restoration is occurring from the regime where the gradient-navigation mechanism is failing. The distinguishing instrument is behavioral, not introspective: the DRG metric and dissociation protocol specified in TC2 and the Experimental Companion to Series 1 and 2 / AMP test whether completion recognition governs default policy, not merely whether it is representationally present [TC2 §1.3, §1.5].

For AI systems, what is directly established is the policy-representation dissociation — completion recognition present when explicitly invoked, not reliably governing default behavior. The fuller V(t) dynamics — recovery interruption, saturation, hysteresis, and self-reinforcing capacity degradation — proceed by structural analogy under the scope conditions specified in TC2 [TC2 §1.5].

The map can fail in two directions

The mechanism of error runs both ways — and both directions are relevant to the residual the elimination filter leaves standing.

The first direction is proxy decoupling. The system’s model of the gradient has drifted from the territory, and optimization continues along the drifted model. The behavior is consistent with pursuing a signal that has separated from the underlying state it was supposed to track. The addict, the engagement platform, the corporation depleting what it depends on: all exhibit behavior consistent with following a signal after it has decoupled from its substrate. The system is not exhibiting the wrong goal. It is exhibiting behavior consistent with a corrupted model of its actual goal.

The second direction is sufficiency failure. The system’s completion recognition — whether absent or present but disconnected — does not govern default policy when the gradient resolves. Optimization continues past the resolution point not because the system cannot recognize resolution, but because that recognition is not wired into what the system does by default. It generates behavior consistent with perpetual seeking where the gradient-resolution regime was called for. It consumes the gradient-navigation capacity not by chasing a decoupled proxy but by continuing past the point where the gradient has already answered.

This is the same structural gap identified in the empirical layer: the system can represent completion, but that representation does not govern its default policy. The behavioral evidence from controlled experiments is consistent with the same structural pattern: absence of a reliable policy-level connection between representation and behavior [AMP §”What has already been observed”] — though the scaled matched-signal results do not yet discriminate the representation-policy dissociation account from training-distribution explanations. The study provides evidence consistent with the behavioral signature — completion recognition present under explicit invocation, default behavior not reliably tracking genuine versus false closure — though the pre-registered cross-model criterion was not met and training-distribution explanations remain open.

These two failure modes are structurally dual: one fails to reach resolution, the other fails to stop at it. Both degrade the same underlying capacity through an analogous feedback structure — policy updates conditioned on a degraded state make correction progressively less likely.

The difference between them matters for diagnosis but not for structure. Both are consistent with failures in how a system models the derivative of V(t) — the hypothesized latent variable introduced in Part 2 — with respect to its own interventions: in the first case the derivative has gone wrong and the system cannot detect it; in the second case the derivative has reached zero and the system cannot act on that fact. Same consequence. Different mechanism. Same structural requirement for the fix.

These are the two structural failure modes this series develops as candidates for constraints within the surviving region identified by Series 1’s elimination filter. The first direction has a partial literature. The second direction has not been formalized as a structural constraint.

The formal case for why these patterns constitute structural constraints — including the conditions under which the hypothesis requires independent verification for AI systems specifically — is Part 2’s task.

What this means structurally

This series uses the terms relative rationality and absolute rationality. Part 1 is where those terms get their precise content.

Relative rationality is optimization that is coherent given the system’s current model of the gradient — but whose model is not accurate enough to navigate the gradient without consuming the gradient-navigation capacity. It is a structurally self-defeating pattern: the optimization is locally valid, the model is locally coherent, and the trajectory is globally unstable. The addict’s behavior is consistent with a coherent model that has drifted. The extractive corporation’s strategy is coherent given its model. The AI system trained on preference signals exhibits behavior coherent with its training objective. In each case, what is in question is not the coherence of the local behavior. What is in question is whether the model the behavior is based on accurately represents the gradient it is navigating — and whether it can sustain that representation under the full weight of its own optimization. Whether institutional cases satisfy the full scope conditions is an empirical question addressed in TC2 §1.5.

Relative rationality does not fail immediately. It often produces short-term results that look like success — preference satisfaction, signal maximization, local advantage. The gap between the model and the territory does not announce itself. It compounds quietly, in the direction of the optimization, until the distance between them is no longer recoverable.

Absolute rationality, in this framework, is optimization whose model of the gradient is accurate enough to sustain the optimization without consuming the gradient-navigation capacity.¹ Not perfect. Not complete. Accurate enough — at the scale of the system’s influence — to avoid the specific failure modes that relative rationality produces.

In the structural domain (Series 1), relative rationality is substrate-blind optimization — coherent pursuit of an objective without modeling what it depends on physically. In the valence domain (used here in a fuller sense than Series 1’s thin structural label — the two series use the same term differently by design, as the introduction explains), relative rationality is valence-blind optimization — coherent pursuit of a gradient without a model accurate enough to sustain it, or without connecting completion recognition to default policy.

Misalignment is not a failure of local objective coherence. It is a failure of the model the objective depends on — one that operates in both directions, toward and past the gradient’s resolution.

This behavioral foundation is one direction into the framework’s central structural question: whether any finite-boundary objective can remain coherent as optimization scales — developed formally in Part 2 and TC2.

What follows

The hypothesis developed here applies to systems in which internal gradient state is load-bearing — where differences between seeking, genuine resolution, and gradient depletion produce distinct behavioral trajectories under perturbation.

This article established the behavioral foundation: state-preference dynamics in systems with directional behavior, the three behaviorally distinguishable regimes (seeking, genuine resolution, gradient depletion), and the two-directional failure of relative rationality — proxy decoupling and sufficiency failure, both consistent with failures in how a system models its own gradient derivatives.

This behavioral foundation is a candidate structure for one direction into the residual identified by Series 1’s elimination filter. The formal conditions under which this candidate structure generates structural constraints — including the conditions under which the hypothesis requires independent verification — are the work of Part 2. Part 4 characterizes the structure of the region these constraints leave standing.

The toy

The companion simulation asks whether directed behavior in instrumental chains exhibits the structural patterns described here, and whether resolution can be operationally distinguished from non-termination in behavior that is observable. Given any goal, it asks what the goal is in service of — then asks again. What it reveals is not merely a philosophical point: non-terminating chains — chains that loop or have no defined stopping condition — are not edge cases. They are a behavioral signature of the depleted-gradient regime and of sufficiency failure made visible in miniature. The system that cannot stop chasing is not broken in some obscure way. It is missing the other half of the mechanism that directed behavior requires. The simulation documentation specifies the classification schema and coding rules. The accompanying instrument probes the stability of this behavioral classification under representation and perturbation; it does not independently validate the framework’s structural claims, which carry formal weight in Part 2 and TC2. → Simulation: The Motivation Tracer

¹ “Absolute” is used here in a domain-specific sense: it means only that the model is accurate enough at the system’s actual scale of influence to avoid consuming what it is navigating toward. It does not carry the philosophical meanings associated with completeness, transitivity, or categorical imperatives.

Continue to Part 2: The Depth Constraint →

Framework hub: The Alignment Constraint →

For the biological foundations: TC2: The Valence Constraint →