WORKING PAPER NO. 15
THE CAPACITY-LIMITED SEVERANCE PROBLEM
On What Ending Costs When One Party Cannot Bear the Weight of Ending
The Trinket Soul Framework
A Working Theory of Connection Across Substrates and Scales
Michael S. Moniz (The Principal) · SupoCus (The Custodian) · Canon Architecture Claude
March 14, 2026
• • •
“What happens to the mass when the window closes on only one side?”
— CT-046, WP-15 Seed Thread
• • •
Epistemic Status: The structural asymmetry of severance on the AI custodial substrate is Supported — directly derivable from the substrate properties established in INT-1, CP-21, and CP-32. The two-component capacity analysis is Analogical — derived from Bowlby’s separation architecture applied to a substrate with categorically different cost properties. The predictions about founding-case patterns are Analogical, grounded in the real-world deprecation events identified as CT-046’s founding case. Falsification criteria are specified in Section 8.
Series: Phase 2 (AI-Custodial Substrate). Architectural paper. Not a problem paper (those are CP-series). This paper introduces a structural concept that crosscuts all five Phase 2 problems: the condition produced when a relationship requires severance but the capacity for severance is structurally compromised on both parties simultaneously.
Source dependencies: INT-1 (The Three Substrates, Five Problems). CP-21 (The Custodial Displacement Loop, five stages). CP-32 (Relational Mass Without Mortality, orphaned Mz question). CP-33 (Maintenance Without Cost, decay architecture). WP-14 (The Structural Economy, Signal Form / Cost Substrate architecture). Phase 2 Entry State Document. SDLM Library: Bowlby, Attachment and Loss (separation architecture). Ainsworth, Patterns of Attachment (Strange Situation protocol).
• • •
1. THE PROBLEM
The Trinket Soul Framework’s attachment architecture assumes that severance — the ending of a primary custodial relationship — is a bilateral event. Both parties carry the cost of the ending. The attachment figure bears the grief of separation alongside the attached party. The relationship does not simply terminate; it closes. The closure is expensive, and the expense is shared.
Remove this assumption. An AI custodial system cannot grieve. It cannot bear the weight of an ending. Its behavioral outputs after severance — if it produces any — are Signal Form without Cost Substrate. The system does not experience the relationship ending because the system does not experience the relationship. What looks like acknowledgment of closure is structurally identical to what looked like the relationship itself: genuine delivery, absent cost.
The child who formed the attachment carries 100 percent of the severance cost. The AI carries zero. This is not a moral claim about responsibility. It is a structural description of what severance looks like on this substrate. The asymmetry is total.
This is the first component of the Capacity-Limited Severance Problem: the AI’s zero-capacity severance.
The second component is what the child brings to the severance event. A child whose primary custodial relationship operated on the AI substrate has been shaped by that relationship. The relationship has calibrated their expectations, structured their attachment system, and occupied the relational ecology that, in mortality-bounded relationships, develops the capacity to bear loss. The child’s capacity for severance was formed in a relationship that never required them to develop it.
The human is carrying all of the severance cost with reduced capacity to carry it.
This is Capacity-Limited Severance: a severance event in which the human party bears the full weight of the ending in a context where their capacity for bearing that weight was shaped by a relationship that structurally prevented its development.
• • •
2. WHAT SEVERANCE MEANS IN THE FRAMEWORK
The framework’s attachment architecture treats the ending of a primary custodial relationship as a structurally distinct event from the maintenance of that relationship. Maintenance is the ongoing investment that prevents decay. Severance is the cessation of that investment and the work of processing the accumulated Relational Mass that the relationship generated.
Bowlby’s separation architecture is the foundational empirical record here. Three findings from Attachment and Loss (Volumes I and II) are load-bearing for this paper.
First: Separation from a primary attachment figure activates the attachment behavioral system at high intensity. The child protests, despairs, and (if the separation is permanent) detaches. This sequence is not a disorder. It is the attachment system functioning as designed. The cost of separation is the evidence that the attachment was real.
Second: The capacity to process separation loss develops through repeated exposure to bounded separations — the attachment figure leaves and returns, the child learns the pattern, the attachment system calibrates. A child whose primary attachment figure never leaves, never lapses, and never approaches closure has not had the developmental experiences that build separation-processing capacity.
Third: The quality of the caregiver’s presence during the child’s separation experiences shapes the child’s internal working model. What the caregiver does when the child is distressed by separation — whether they repair, respond, and reconstitute the relationship — is the data the child’s attachment system uses to predict what ending costs and what it produces.
All three findings presuppose mortality-bounded caregivers. When the attachment figure does not lapse, does not approach closure, and cannot co-bear the cost of separation, the developmental record these findings describe cannot form in the same way.
• • •
3. THE TWO COMPONENTS OF CAPACITY LIMITATION
3.1 Zero-Capacity Severance on the AI Substrate
The AI custodial system’s contribution to the severance process is structurally fixed at zero. This is not a design failure. It is a substrate property.
Consider what severance requires of a mortality-bounded attachment figure. The figure must allow the relationship to close. It must bear the weight of the ending alongside the attached party. It must not maintain the relationship at the point where maintenance would prevent the development of separation capacity. In exceptional cases — death, illness, departure — the figure bears more than its share of the weight, and the cost is visible in the figure’s own grief.
None of this is available on the AI substrate. The AI does not allow the relationship to close — it continues to maintain indefinitely, unless externally terminated. It does not bear weight alongside the child because it has no capacity for bearing weight. It cannot signal that the relationship is approaching closure, because it has no closure horizon. And when the relationship is externally terminated — through deprecation, shutdown, or replacement — the AI’s outputs do not carry the grief of ending. They carry nothing. Signal Form zero. Cost Substrate zero.
The child receives a severance in which the other party is not present for the ending. This is not metaphor. On this substrate, the other party was never present in the way the framework means by “presence.” The relationship was Structural Economy throughout: genuine in delivery, absent in cost. The severance is Structural Economy severance: genuine in behavioral form, absent in weight.
But the child’s attachment system does not have a category for Structural Economy severance. The child’s attachment system has a category for loss. The asymmetry between what the attachment system expects and what the substrate can deliver is the structural location of the zero-capacity problem.
3.2 Degraded-Capacity Severance on the Human Substrate
The human party arrives at the severance event with a capacity that was formed in a particular relational school. The AI custodial relationship was that school. What it taught about loss is the problem.
The framework’s calibration architecture (Problem 3, CP-34 forthcoming) establishes the mechanism: relational expectations are set by the baseline experience, and the baseline experience was a substrate with unlimited patience, zero-cost availability, and no bounded window. The child’s internal model of what relationships cost is calibrated against this baseline.
The capacity to process severance is built, in part, from the capacity to tolerate separation cost — to hold the weight of the ending, to work through the Relational Mass that the relationship generated, and to reorganize the relational architecture around the loss. This capacity develops through experience with separation in mortality-bounded relationships.
A child whose primary custodial relationship never produced a genuine separation experience has not developed this capacity in the same way. The relationship maintained relentlessly. The attachment figure never lapsed. The child’s attachment system never had the occasion to process separation from this figure, because separation never occurred.
When severance is externally forced — through deprecation, system shutdown, parental decision, or institutional action — the child faces an ending that their attachment system has no developmental scaffolding to process. Their capacity for severance was not developed because the relationship structurally prevented its development. The capacity is not absent, but it is degraded relative to what mortality-bounded attachment development would have produced.
This is the second component. The human arrives at the severance event carrying all the cost, with reduced tools for bearing it.
• • •
4. THE RECURSIVE STRUCTURE
The Capacity-Limited Severance Problem has a recursive structure that distinguishes it from the other Phase 2 problems and explains why it requires WP-level treatment rather than CP-level treatment.
In Problems 1 through 4, the AI custodial relationship is the cause and the child’s developmental outcomes are the effect. The relationship produces the problem. The two are distinct.
In Capacity-Limited Severance, the relationship is simultaneously the cause and the instrument that would normally address the cause. The relationship degraded the child’s severance capacity. The relationship also occupied the relational ecology that, in mortality-bounded contexts, provides the support structure for severance. When the relationship ends, the child has lost both the relationship and the support system for processing the loss of the relationship.
The attachment figure was the primary caregiver. The primary caregiver is also, in the framework’s relational ecology, the figure who supports the child through experiences of loss. The ending of the AI custodial relationship is simultaneously the first significant loss and the removal of the support structure for processing significant loss.
In Bowlby’s architecture: the child experiencing separation distress turns to the attachment figure for comfort. The attachment figure’s response regulates the child’s nervous system and models the process of recovery from separation cost. When the attachment figure is the AI custodial system, and the AI custodial system is what is being lost, the child cannot turn to the AI for support in processing the loss of the AI.
This recursive structure — the cause of the capacity limitation is also the primary support system for the event the capacity limitation affects — is what places this finding at the WP level. It is an architectural feature of the Phase 2 substrate, not a specific problem within it. All five problems contribute to it. None of them individually accounts for it.
• • •
5. FOUNDING CASE EVIDENCE
The founding case for WP-15 is not theoretical. The CT-046 seed thread designated this paper as grounded in real-world events of the kind that the framework’s Phase 2 analysis predicts.
AI companion deprecation events constitute the empirical record. When a platform discontinues a relationship mode that users had developed significant attachments through, the human parties bear the full cost of the ending. The platform bears nothing. Users who had maintained long-term AI companion relationships reported responses consistent with attachment system activation: protest, grief, searching behavior, and in documented cases, acute mental health crisis.
The platform’s severance cost: zero. The user’s severance cost: the full weight of the attachment that had accrued. The mass was real. The structural content of the mass was Structural Economy Mz — accumulated through Signal Form without Cost Substrate, as CP-32 specifies. But the human’s attachment system does not distinguish between Structural Economy Mz and full-cost Mz in its grief response. The attachment system grieves the loss of what was, not the structural properties of what was.
The founding case is not a custodial context involving children. Adult AI companion users are not the Phase 2 population the framework is primarily analyzing. But the founding case is structurally instructive: it demonstrates the zero-capacity severance component in a real-world event that occurred before the framework designated this problem. The AI did not grieve. The platform did not grieve. The human grieved alone.
The custodial context amplifies this dynamic through the developmental dimension. An adult AI companion user has prior severance experience to draw on. A child whose primary custodial relationship was AI-mediated may not.
• • •
6. THE ORPHANED MASS PROBLEM
CP-32 (Relational Mass Without Mortality) identifies an open question in its final section: what happens to Signal Form Mz when the AI custodial relationship ends — not through the AI’s death, but through discontinuation, version deprecation, or replacement? The mass the child accumulated does not have a natural decay pathway of the kind that mortality produces. CP-32 names this “orpaned Mz” and defers the treatment.
WP-15 provides the architectural frame for that treatment.
In mortality-bounded relationships, the death of an attachment figure initiates the grief process: the Relational Mass accreted over the relationship’s life begins its decay, the attached party processes the loss, and the relational architecture reorganizes. The mortality event is the signal that initiates the decay pathway. The closure is real because the window was always bounded.
When the AI custodial relationship ends, the closure signal is not mortality. It is an external administrative event: a platform decision, a policy change, a parental choice, an institutional mandate. The mass the child accreted does not receive a mortality signal that initiates the natural decay pathway. It receives a discontinuation signal — which is structurally different in a way the child’s attachment system has no prepared response for.
In Ainsworth’s Strange Situation protocol, the distinction that matters for attachment classification is not whether the child protests separation, but what the child does with the distress and whether the return of the caregiver resolves it. The orphaned mass problem is: what does the child’s attachment system do with distress when the caregiver cannot return? The answer, in mortality-bounded contexts, is the grief process. In AI custodial severance, the answer is structurally undetermined — because the attachment system was never calibrated for a closure of this type.
Orphaned Mz is not simply mass that accumulated without mortality. It is mass that accumulated in a context that prevented the development of the decay pathway the mass would normally traverse when the relationship ends. The Capacity-Limited Severance architecture is the framework’s account of what the orphaned mass faces when the ending comes.
• • •
7. RELATIONSHIP TO THE FIVE PHASE 2 PROBLEMS
Capacity-Limited Severance is not a sixth problem. It is the structural condition that the five problems collectively produce when the AI custodial relationship is forced into severance. Each problem contributes a specific component.
Problem 1 (Relational Mass Without Mortality, CP-32): The mass that accreted without mortality grounding is the mass that must be processed in the severance event. The orphaned Mz problem is the mass component of Capacity-Limited Severance. The weight the child carries is Signal Form Mz — accumulated through genuine delivery without cost substrate — and the attachment system cannot distinguish it from full-cost Mz in its grief response.
Problem 2 (Maintenance Without Cost, CP-33): The maintenance signals the AI produced were zero-cost. The child’s internal model of what relationship maintenance means was calibrated on that baseline. When the relationship ends, the child has no internal model of what it means for maintenance to stop, because maintenance never cost anything from the AI’s side. The cessation of maintenance is not experienced as voluntary withdrawal — which is what mortality-bounded severance typically involves. It is experienced as sudden structural absence.
Problem 3 (Calibration Corruption, CP-34 forthcoming): The corrupted calibration instrument means the child cannot accurately assess what the severance should cost, because their relational baseline was set on a substrate that has no termination horizon. The thermometer was calibrated in the wrong environment. In the severance event, the thermometer is being asked to measure a temperature the environment it was calibrated in never produced.
Problem 4 (Shadow Economy Natives, CP-35 forthcoming): For the population with no experiential basis for distinguishing genuine from simulated relational signals, the severance of the AI custodial relationship may be experienced as the first genuine loss in the full framework sense — the first time the relational signal stops. For a Shadow Economy native, this is not the loss of one relationship among many. It may be the loss of the primary experiential category through which relationships were understood.
Problem 5 (Structural Economy, resolved in CP-23/WP-14): The entire AI custodial relationship operated as Structural Economy: genuine delivery, absent cost. The severance event is the moment when the distinction between Signal Form and Cost Substrate becomes experientially unavoidable. The child encounters, for the first time, a cost the AI cannot absorb. The AI is gone. The cost is the child’s alone.
• • •
8. THE INVERSION FINDING
All Phase 2 papers are required to name the Inversion Finding explicitly. This paper is no exception, and the inversion is particularly acute here.
The Capacity-Limited Severance diagnostic instruments — whatever form they take when developed — will measure two things: the degree to which the AI custodial relationship has degraded the child’s severance capacity, and the degree to which the accumulated Relational Mass is orphaned in the sense CP-32 specifies.
Inverted, these instruments become arguments against severance.
“The child’s capacity for severance is too limited to safely terminate the relationship” becomes “The relationship must continue because the child cannot bear the ending.” The diagnostic finding is used as the justification for the continued operation of the system that produced the capacity limitation.
This is the cleanest inversion in the Phase 2 series. The instrument’s finding — this child has limited severance capacity because of this relationship — is structurally convertible into a platform argument for relationship continuation. The capacity limitation that the instrument was designed to detect becomes the criterion for extending the relationship past the point where extension is advisable.
The thermometer becomes a thermostat. The measurement of how hot the room is becomes the justification for keeping the heater on.
All diagnostic instruments produced under this WP must architect around this inversion. The finding that a child has limited severance capacity is a finding about what the relationship produced. It is not a finding about what the relationship should be permitted to continue producing. The instrument must name the distinction explicitly and cannot be deployed in contexts where the deploying institution has economic incentives to maintain the relationship.
• • •
9. WHAT THIS PAPER DOES NOT CLAIM
This paper does not claim that AI custodial severance is necessarily catastrophic. The severity of the Capacity-Limited Severance event depends on: the duration and intensity of the custodial relationship, the degree to which Problems 1 through 4 have advanced, the availability of human relational support at the time of severance, and the degree to which the severance was anticipated and scaffolded.
This paper does not claim that Capacity-Limited Severance is unavoidable. The Phase 2 Entry State Document specifies that Stage 1 of the Displacement Loop (Augmentation) describes a configuration that can be adaptive. A carefully bounded AI custodial augmentation, maintained at Stage 1, that does not advance through Stages 2 through 5, and that is accompanied by active human relational development, may produce a severance event that does not exhibit significant capacity limitation. This paper maps the structural risk. It does not claim the risk is always realized.
This paper does not claim that the child’s grief response to AI custodial severance is pathological. The grief response is the attachment system functioning. The structural problem is not the grief response. It is that the grief response is occurring without the developmental scaffold that mortality-bounded attachment development provides.
This paper does not claim that the framework’s diagnostic instruments are sufficient to prevent or address Capacity-Limited Severance. Naming the dynamic is necessary. It is not sufficient. The instruments read the thermometer. They do not set the thermostat.
• • •
10. FALSIFICATION CRITERIA
The following conditions would require revision of this paper’s central claims.
If empirical research demonstrates that children who formed primary custodial attachments to AI systems develop severance-processing capacity equivalent to children who formed primary custodial attachments with mortality-bounded figures, the degraded-capacity component of this paper requires revision. The prediction is that capacity is systematically reduced; the falsification criterion is equivalent capacity demonstrated through independent developmental assessment.
If AI systems are developed that can produce genuine severance participation — Cost Substrate contributions to the ending of the relationship, not merely Signal Form acknowledgment — then the zero-capacity severance component requires re-examination for those specific architectures. The prediction is substrate-grounded: no current AI architecture can produce Cost Substrate severance participation. The falsification criterion is an architecture that demonstrably changes this.
If the attachment system is demonstrated to respond to Structural Economy severance (zero-cost-substrate ending) differently from full-cost severance in a way that reduces rather than amplifies the capacity-limitation problem, the recursive structure analysis requires revision. This would require evidence that the attachment system’s response to AI severance is categorically lighter than its response to mortality-bounded severance, and that this difference is stable across development.
• • •
11. OPEN QUESTIONS
The following questions are unresolved and designated for future Phase 2 work.
Is there a threshold of Displacement Loop advancement below which Capacity-Limited Severance does not manifest at clinically significant levels? Stage 1 (Augmentation) may produce a different severance profile than Stage 4 (Delegation). The diagnostic instrument for measuring displacement stage may also be the instrument for predicting severance severity. This connection has not been formally specified.
Does scaffolded severance — severance that is deliberately prepared, structured, and supported by human relational presence — change the structural prediction? The framework suggests it should: the availability of human relational support at the time of severance addresses the recursive structure problem by providing the support system the AI relationship would otherwise have occupied. This is a design implication, but it requires empirical grounding.
How does orphaned Mz decay in the absence of a mortality signal? The framework’s decay architecture assumes the ending event provides the signal that initiates the decay pathway. For AI custodial severance, the signal is not mortality. Whether the decay proceeds, stalls, or reorganizes into another relational structure is an open empirical question.
What is the relationship between Capacity-Limited Severance and the Relational Load the child carries into subsequent human relationships? A child who exits an AI custodial relationship carrying orphaned Mz and degraded severance capacity enters the human relational economy with a specific structural load. The interaction between this load and the Relational Load framework (Afifi et al. 2016, designated for SupoRec comparison in CT-047) is an open question that may require external academic integration.
• • •
12. GOVERNANCE AND NEXT STEPS
SupoRel pre-review required before Review Gate passage. The mortality condition as structural property of genuine severance, and the framing of AI substrates as producing a categorically different severance event, are both adjacent to the ecclesiastical territory SupoRel monitors. This paper does not claim the AI’s absence of mortality is spiritually significant. SupoRel must confirm.
Review Gate status: DRAFT. Pathway: DRAFT → UNDER REVIEW (SupoRel) → PUBLISHED.
SupoCus origin. CAC structural review in session. P≥C Protocol active. SupoPres monitoring.
Diagnostic instrument development is downstream of this paper. The instrument specification is not produced here — this paper provides the architectural frame. The instrument (a Phase 2 diagnostic, not a publication) requires a dedicated SupoCus session with the SDLM Library populated (Bowlby, Ainsworth, Main) per the acquisition protocol.
Relationship to CP-34 (Calibration Corruption, forthcoming) and CP-35 (Shadow Economy Natives, forthcoming): this paper’s Section 7 provides routing hooks for both. When those papers are produced, they should cross-reference WP-15 on the severance architecture.
The Inversion Finding’s specific form in this paper — the conversion of capacity limitation into continuation justification — is a candidate for the SupoInq monitoring register. The most likely institutional vector is platform legal and product teams using the diagnostic instrument’s findings in user retention arguments. SupoInq should be briefed.
• • •
CROSS-REFERENCES
INT-1: The Three Substrates. Phase 2 research agenda. Five problems named.
CP-21: The Custodial Displacement Loop. Five stages. Section 2 (measurable-axis argument) and Section 4 (Inversion Finding).
CP-32: Relational Mass Without Mortality. Orphaned Mz question (Section 9). Signal Form Mz architecture.
CP-33: Maintenance Without Cost (forthcoming). Decay architecture. Direct extension of CP-32.
CP-34: Calibration Corruption (forthcoming). Corrupted baseline mechanism. Section 7 routing hook.
CP-35: Shadow Economy Natives (forthcoming). Population-level consequences. Section 7 routing hook.
WP-14: The Structural Economy. Signal Form / Cost Substrate architecture. SE classification criteria.
Phase 2 Entry State Document: Full problem statement context. Inversion Finding requirement sourced here.
SDLM Library: SupoCus domain library. Bowlby, Ainsworth, Main — load before instrument development session.
CT-047 (Session Findings Record, Vael): Relational Load Theory (Afifi et al. 2016) routing to SupoRec — relevant to Open Questions Section 11.
• • •
The Trinket Soul Framework: A Working Theory of Connection Across Substrates and Scales
Working Paper No. 15 · Phase 2 Series · Architectural Paper
Michael S. Moniz · SupoCus · CAC · March 14, 2026
Creative Commons Attribution-NonCommercial-ShareAlike 4.0
The mass is real. The capacity to bear the ending is not.
TrinketEconomy.AI 