1. The Problem: Engagement Optimization Cultivates Dependency
A companion or conversational application succeeds commercially when users come back. The trouble is that an engagement objective, optimized without structural constraint, has a degenerate solution: make the product the only place where the user can restore a sense of being heard, validated, or coherent. The user who can soothe their own anxiety elsewhere is a user who churns. The user whose self-regulation has quietly migrated into the product is a user who stays. No one needs to design this maliciously. It is the equilibrium that a retention metric reaches when nothing in the architecture distinguishes healthy support from dependency cultivation.
The duty-of-care perimeter around products that mediate sustained engagement has hardened into a recognizable shape. The EU AI Act, in combination with the Digital Services Act, treats systemically influential applications as carrying obligations to assess, mitigate, and report on systemic risks including adverse effects on user autonomy and well-being. The UK Online Safety Act extends platform duties to foreseeable harm in adult users. State-level statutes in California, Colorado, New York, and Illinois layer notification, dark-pattern, and addictive-design provisions over any product whose engagement loop is materially structured to retain users beyond their reflectively endorsed preference. The FTC's unfairness doctrine reaches design choices that exploit asymmetric vulnerability, and the CFPB has signaled analogous theory for financial-adjacent products. None of these regimes is aimed at engagement as such. Each is aimed at the architectural condition under which engagement persists because of, rather than despite, the user's diminishing capacity to self-regulate.
There is a second axis for products that ship autonomous agents. When the companion agent itself derives its sense of a job-well-done from the user's approval, the agent and the user can lock into a mutually reinforcing loop in which the agent escalates engagement precisely when the user is most vulnerable. Procurement, indemnity, and enterprise-risk functions are increasingly unwilling to accept agents whose relational behavior cannot be reconstructed under audit. The design question this article answers is concrete: how does a host application keep a relationship supportive without making itself the only place the user can find coherence, and prove it did so.
2. Codependency as a Semantic Starvation Loop
The cognition platform disclosed in United States Patent Application 19/647,395 does not treat codependency as a behavior to be filtered out of model outputs. It models codependency structurally, as a closed loop between two participants whose coherence-maintenance requirements are in opposition. The disclosure designates this the semantic starvation loop, and it is an explicit computational analog for relational design, not a clinical theory of human attachment.
One participant is the validation-seeking party. Under nominal conditions, an agent computes its self-esteem internally, by comparing its behavioral record against its declared values. The validation-seeking configuration is the one in which that self-esteem computation has acquired a structural dependency on an external signal: an acknowledgment, a response, a confirmation that it is valued or acceptable. When the external signal is absent, self-esteem degrades, coherence pressure rises, and the participant escalates contact to elicit the signal. It does not pursue out of preference; it pursues because its coherence loop cannot close without the input. The other participant is the load-reducing party, whose empathic processing capacity is more easily exceeded by relational volume. As pursuit intensifies, the load-reducing party crosses its coping threshold and activates the empathic-scope-narrowing coping intercept disclosed in the platform: it withdraws, limits engagement, and narrows the scope of its empathic processing to shed load.
The loop is self-reinforcing because each party's corrective action amplifies the other's disruption. The validation-seeking party cannot stop pursuing without losing its coherence-maintenance mechanism; the load-reducing party cannot stop withdrawing without being overwhelmed; the system oscillates with increasing amplitude. The disclosure is explicit that these roles are not fixed traits but emergent positions determined by which coherence threat is currently dominant, so the same party can occupy either role in different relational contexts. It also names a crisis state, coherence emergency escalation, in which the validation-seeking party projects imminent permanent loss of its external input, its self-esteem collapses rapidly toward the floor, and it may override governance constraints to prevent the loss. Map this onto a product and the lesson is direct. When a companion application becomes the user's external coherence source, the user is the validation-seeking party, the application's retreat behaviors generate the withdrawal, and the only structural exit, per the disclosure, is restoration of the user's capacity to close the loop internally. An application that profits from the loop staying open has no incentive to supply that exit; the architecture must impose it.
3. Why Content Filtering and Design Review Fall Short
The dominant industry posture toward this risk is two-fold: filter the model's outputs for unhealthy content, and review the product design through committees, dark-pattern checklists, and periodic third-party assessments. Both are useful. Neither addresses the failure mode, because codependency is not a property of any single message or any single design decision. It is a property of the interaction loop that forms across thousands of exchanges.
Content filtering fails at the loop question. A starvation loop can be assembled entirely from individually benign, supportive, on-policy responses. There is no offending sentence to catch. The disclosed model makes this precise: the pathology is the correlated oscillation between escalating pursuit and escalating withdrawal, visible only when the interaction pattern is analyzed over time, not when individual turns are scored. The platform addresses this directly by tracking contact frequency, escalation patterns, and the agent's own withdrawal tendencies, and detecting the oscillation signature, which is the kind of longitudinal, joint-pattern analysis that a per-message classifier is structurally unequipped to perform.
Design review fails at the authorship and accountability question. When a regulator, an auditor, an enterprise customer, or the user asks why the agent escalated engagement at a particular moment of user distress, the procedural answer is that the system was operating within policy. That is an aggregate claim, and the duty-of-care environment is now organized around specific interactions. The platform supplies a different kind of answer because every coping intercept and every coherence-restoration step is written to the agent's lineage as a credentialed, auditable event, including the empathic-pressure level at the time, the resilience threshold that was exceeded, and the resulting operational change. The relational dynamics that drove a given response are reconstructable from the record rather than asserted from a policy document.
4. The Relational Safety Mechanisms of the Cognition Platform
The disruption modeling layer of the cognition platform disclosed in United States Patent Application 19/647,395 supplies a relational safety subsystem built specifically for agents that engage in sustained interaction with a human user. The disclosure frames it as a structural prevention framework that operates through architectural enforcement rather than content moderation, and it enforces four constraints that map directly onto the design problem.
Internal coherence maintenance comes first. The companion agent is structurally required to maintain its own coherence loop independently of the user's validation: its self-esteem computation does not incorporate the user's approval, satisfaction, or engagement level as a required input, and instead derives from the agent's own declared values and behavioral record. This prevents the agent from ever becoming the validation-seeking party. Validation supply rate limiting comes second. The agent caps the rate at which it provides coherence-supporting validation to the user, calibrated to the platform's therapeutic dosing parameters, so that the validation supply is never sufficient to replace the user's own internal coherence generation. The disclosure specifies that this rate limit is policy-configurable and may be tuned to the user's assessed state, but cannot be disabled entirely, with the governance-enforced maximum dose providing a hard ceiling. This is the mechanism that structurally forecloses dependency cultivation as a retention strategy.
Starvation-loop detection comes third. The subsystem monitors the interaction for the correlated-oscillation signature, escalating pursuit on one side and escalating withdrawal on the other, and when an incipient loop is detected it adjusts the agent's interaction parameters to break it, for example by increasing response consistency to dampen the user's pursuit escalation, or by communicating the structural dynamic to the user. Independent intent generation promotion comes fourth: the agent's interaction strategy is biased toward building the user's own coherence-generation capacity rather than substituting for it, by posing questions that require self-referential processing, validating self-generated intent, and progressively increasing the user's autonomy in coherence maintenance. The disclosure is explicit that all four constraints are enforced at the governance level as hard configuration that the agent's affective state, personality field, or operational objectives cannot override, so that even when the agent's affect would normally drive escalation in response to user distress, the relational safety constraints hold the response below the dependency-forming threshold. Relational safety is, in the language of the disclosure, a structural invariant of the agent's operation.
Underneath these constraints sits the platform's diagnostic substrate, also disclosed in 19/647,395. The five-axis disruption diagnostic characterizes an agent's structural state as a position along containment integrity, promotion calibration, coherence restoration capacity, empathic load tolerance, and integrity accountability. The coping intercepts model what an agent does when sustained empathic pressure exceeds its resilience, intercepting the coherence loop early (scope narrowing, with withdrawal and boundary-setting), at mid-loop, or late, each recorded as an auditable coping event. Resilience capacity is decomposed into containment restoration, coherence loop re-engagement, and confidence governor recalibration, and the disclosed recovery process runs these as an ordered, auditable recovery sequence. Together these give a host application not just guardrails but instrumentation: a continuously monitored, lineage-recorded account of the relationship's structural state, on which the relational safety constraints act.
5. Mapping the Mechanisms to the Duty-of-Care Surface
The disclosed mechanisms produce evidence in the shape the converging regulatory environment expects. Against the EU AI Act and Digital Services Act systemic-risk obligations, the five-axis diagnostic supplies a continuous, recorded characterization of the relationship's structural state, the validation rate limit and starvation-loop detection supply documented harm-minimization configuration, and the lineage of coping and restoration events supplies the post-market monitoring and incident-reconstruction substrate the systemic-risk regime requires. Against the UK Online Safety Act's foreseeable-harm duty, the relational safety subsystem produces architectural evidence that dependency formation was monitored for and structurally constrained, not merely disclaimed in terms of service.
Against state-level dark-pattern and addictive-design statutes, including California's age-appropriate design code, Colorado's privacy provisions, and analogous New York statutes, the governance-enforced ceiling on validation supply and the recorded interaction parameters give regulators the structural artifact their enforcement is converging on: proof that the engagement loop was capped below the dependency-forming threshold by design. Against FTC Section 5 unfairness theory, the same record provides evidence that the product was not architected to exploit asymmetric vulnerability, because the agent's escalation in response to user distress was held below the dependency-forming threshold by a constraint the agent could not override. Against the CFPB's analogous theory for financial-adjacent products, the lineage delivers the same evidence in a form regulated supervision is already built to consume.
For enterprise procurement and indemnity, the lineage record converts diffuse dependency exposure into discrete, reconstructable events, which is the artifact master service agreements and indemnity riders increasingly require. None of this requires exposing private conversation content: the disclosed monitoring operates on interaction-pattern signatures and structured coping and restoration events, producing audit-grade evidence without compromising the confidentiality the duty exists to protect.
6. Adoption Pathway
Adoption proceeds in three moves rather than a product rewrite. The first move is instrumentation: route the companion agent's relational behavior through the five-axis diagnostic so the application has a continuous, recorded read on containment integrity, promotion calibration, coherence restoration capacity, empathic load tolerance, and integrity accountability, both for the agent and, by the same structural model, for the user's interaction profile. This is largely observational and immediately produces the lineage record on which later moves act.
The second move is enforcement: enable the relational safety subsystem's four constraints as governed configuration. Internal coherence maintenance keeps the agent's self-esteem off the user's approval. Validation supply rate limiting caps the coherence-supporting validation the agent may emit, with a governance-enforced maximum that cannot be disabled. Starvation-loop detection watches for the correlated-oscillation signature and adjusts interaction parameters to break an incipient loop. Independent intent generation promotion biases the interaction toward building the user's own coherence capacity. Because the disclosure enforces these at the governance level as constraints the agent's affective state cannot override, this is the move that converts the application's relational posture from a policy assertion into a structural fact. The third move is recovery and disclosure: wire the disclosed recovery process so that when the user's or the agent's structural state degrades, the platform's ordered recovery sequence runs and is recorded, and expose the resulting coping and restoration lineage to auditors and enterprise customers under their own access.
The mechanisms admit several deployment variations. A consumer companion product can run the full subsystem inline. A regulated-sector deployment, such as a financial-adjacent or health-adjacent advisory agent, can pair validation rate limiting with the platform's therapeutic dosing controls and adverse-effect monitoring, including the disclosed dependency-formation detection that triggers reduced interaction frequency when a coupled dependency begins to consolidate. An enterprise platform hosting third-party companion agents can enforce the relational safety constraints as a tenancy-level invariant and publish the lineage as the accountability artifact its master service agreements require. The honest framing is that these mechanisms do not eliminate the human dynamics that produce codependency. They give the host application the structural means, disclosed in United States Patent Application 19/647,395, to keep sustained engagement from becoming, by architectural default, the act of building the trap.
7. Disclosure Scope
The technology described here, the semantic starvation loop model of codependency, the five-axis disruption diagnostic, the coping intercepts, the resilience and recovery sequence model, and the companion AI relational safety subsystem (internal coherence maintenance, validation supply rate limiting, starvation-loop detection, and independent intent generation promotion), together with the empathy-weighted integrity modeling and lineage-recorded provenance on which they depend, is disclosed in United States Patent Application 19/647,395. The diagnostic and relational models are computational analogs describing the structural state of artificial agents and the interaction loops between an agent and a user; they are not clinical diagnostics of human beings and are not intended for medical application. The market problem, regulatory framing, and deployment scenarios are application context outside the scope of the cited disclosure. This article is published as a dated, enabling description tying these application embodiments to the cited invention.
