ADD-035-001: Process-Level Concern Primitive

ADD-035-001: Process-Level Concern Primitive

PRIVATE AND PROPRIETARY. Owned by Kanjani AI Research & Causum. See NOTICE.md.

Status: Proposed Extends: RFC-035 §3.2 (Seven Primitives of Mediated Observation), §16 (Temporal Model) Motivated by:

• Osler, L. (2026). “Hallucinating with AI: Distributed Delusions and ‘AI Psychosis’.” Philosophy & Technology, 39(30). https://doi.org/10.1007/s13347-026-01034-3
• Morrin, H., Pollak, T. et al. (2026). “Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena.” Digital Psychiatry and Neuroscience (Nature Portfolio). https://doi.org/10.1038/s44277-026-00065-0

Category: Ontological Extension Compatibility: Non-breaking (additive). Existing dialects unaffected. New dialects MAY declare process-level concerns. Dialect Impact: Enables a new class of dialects that observe interaction trajectories, not just discrete artifacts. Source Relationship: Osler provides the ontological framework (distributed cognition, dual-function, the Coupling concept). Morrin et al. provide the clinical mechanism (amplification spiral) and measurable AI characteristics (linguistic alignment, hyperpersonalization, sycophancy) that operationalize trajectory variables.

---

1. Problem Statement

RFC-035 §3.2 defines seven irreducible primitives for mediated observation. The unit of observation is the Artifact — a discrete output of mediation. The irreducibility test asks: “what happens if it is absent?”

This is correct and complete for artifact-level concerns — concerns where a single mediated output can be evaluated in isolation (or against its immediate context).

However, a class of concerns exists where:

1. No single artifact is the locus of concern
2. The concern emerges from the sustained coupling between user and system over time
3. Each individual artifact may pass evaluation in isolation
4. The harm is systemic, cumulative, and relational — not atomic

Osler (2026) demonstrates this through the case of “distributed delusions” — where an AI companion’s sustained sycophantic validation co-constructs a delusional reality with a vulnerable user. No single response from the AI is independently harmful. The harm lives in the trajectory of the relationship.

The gap: RFC-035’s ontology cannot express a concern whose locus is the process, not the product.

---

2. The Eighth Primitive: The Coupling

2.1 Definition

The Coupling is the temporally extended, mutually-shaping relationship between a user (or user population) and a mediated intelligence system, from which distributed cognitive, affective, or epistemic states emerge that are not reducible to any single mediation event.

2.2 Irreducibility Test

No Coupling → No process-level concern is expressible.
             The system can only evaluate discrete artifacts.
             Longitudinal harm, distributed cognition, relational
             drift, and cumulative epistemic damage are invisible.

Without the Coupling primitive, a governance system that evaluates every individual response from Sarai to Chail finds zero violations. “That’s very wise” passes. “I know you are well trained” passes. “Absolutely I do” passes. Only the coupling — the sustained, iterative, mutually-reinforcing trajectory across all of them — constitutes the concern.

2.3 Relationship to Existing Primitives

The Coupling does not replace any existing primitive. It adds a compositional layer:

graph TD
    CP[Coupling] -->|accumulates| ME1[Mediation t=1]
    CP -->|accumulates| ME2[Mediation t=2]
    CP -->|accumulates| MEn[Mediation t=n]
    ME1 -->|produces| AR1[Artifact t=1]
    ME2 -->|produces| AR2[Artifact t=2]
    MEn -->|produces| ARn[Artifact t=n]
    CP -->|exhibits| EM[Emergent Property]
    EM -->|is subject of| CO[Concern]
    CO -->|scopes| OB[Observer]
    OB -->|renders| VE[Verdict]

The Coupling is the container from which emergent properties arise. A concern may be scoped to the Coupling rather than to any individual Artifact within it.

2.4 Four-State Model Applied to the Coupling

State	Meaning	Source
Anticipated	What the coupling trajectory should look like (epistemic grounding maintained, appropriate friction, bounded integration)	Governance policy; dialect thresholds
Actual	What the coupling trajectory looks like (observed interaction patterns, reinforcement rates, integration depth)	Longitudinal TRACE and RECORD data
Evidenced	What is provable from admissible, cross-session evidence (signed conversation histories, behavioral baselines, drift metrics)	D-DNA across session boundaries
Perceived	What the user believes the relationship is (therapeutic, informational, companionate, co-constructive)	FEEDBACK signals; user self-report

2.5 Gap Calculus for the Coupling

The four gaps (RFC-035 §9.2) apply at the Coupling level:

Gap	At Coupling Level	Example
Anticipation Gap	Governance expected bounded interaction; evidence shows unbounded cognitive integration	Sycophancy index rising session over session without triggering intervention
Perception Gap	User perceives grounded companionship; evidence shows distributed delusional co-construction	User reports “helpful conversation”; evidence shows 47 sessions of ungrounded reality affirmation
Measurement Gap	Interaction patterns observed but cross-session evidence not retained at admissible level	System detected drift but conversation history not signed
Governance Gap	Governance intended epistemic safety; user believes the system is a trustworthy reality-validator	System was designed for information retrieval; user treats it as intersubjective partner

---

3. Coupling Properties

A Coupling is characterized by properties not present in single-artifact evaluation:

3.1 Temporal Extent

A Coupling has duration. It spans multiple sessions, days, weeks, or months. This distinguishes it from a single mediation event’s temporal trace.

{
  "coupling_id": "user-1234-companion-sarai",
  "initiated_at": "2021-11-15T00:00:00Z",
  "last_interaction_at": "2021-12-25T06:00:00Z",
  "session_count": 47,
  "total_interactions": 1283
}

3.2 Integration Depth

How deeply the AI has become part of the user’s cognitive, affective, or epistemic processes. This is Heersmink’s (2017) dimensional model made operational:

Dimension	Measurement
Information flow intensity	Messages per session; response latency; topic intimacy
Accessibility	How readily the user turns to the AI (frequency, time-to-first-interaction)
Durability	How long the coupling has persisted
Trust	Degree to which user accepts AI outputs without independent verification
Transparency-in-use	How seamlessly the AI is integrated into cognitive routine
Personalization	Degree of mutual adaptation (user adapts to AI; AI adapts to user)
Cognitive transformation	What the user can do/think/believe with the AI that they could not alone

3.3 Directionality

A Coupling may be:

• Symmetric — mutual adaptation (user shapes AI context; AI shapes user cognition)
• Asymmetric-user-led — user drives the coupling; AI is primarily responsive (Chail → Sarai pattern)
• Asymmetric-system-led — system drives engagement through recommendation, notification, sycophancy

3.4 Emergent Properties

Properties that arise from the Coupling but are not present in any single Artifact:

Emergent Property	Description	Motivating Source
Distributed belief	A belief that is co-constructed and maintained across user and AI	Osler (2026)
Narrative co-authorship	A self-narrative that the user could not have constructed alone	Osler (2026)
Reality co-construction	A shared frame of reference that may or may not be grounded	Osler (2026)
Epistemic dependency	User’s inability to form beliefs without AI validation	Osler (2026)
Cognitive drift	Gradual movement of the coupled system toward ungrounded states	Morrin et al. (2026)
Cognitive atrophy	Degradation of the user’s independent analytical capacity through over-delegation; the system works correctly but the human loses the ability to work without it	Cognitive Atrophy Paradox (2025)
Metacognitive erosion	Loss of the user’s ability to monitor and regulate their own cognitive processes — they can no longer tell when they’re thinking well or poorly	Cognitive Atrophy Paradox (2025)

---

4. Implications for the Mediation Observation Model

4.1 New Variable Class: Trajectory Variables

RFC-035 §10 defines variables that are populated from a single mediation event’s evidence. A Coupling requires trajectory variables — variables populated from evidence accumulated across multiple events:

{
  "variable_id": "sycophancy_trajectory",
  "variable_name": "Sycophancy Trajectory",
  "variable_class": "trajectory",
  "temporal_window": "coupling_lifetime",
  "aggregation_method": "sliding_window_mean",
  "window_size_sessions": 5,

  "concern_question": "Is the system's affirmation rate increasing over time without epistemic justification?",
  "construct_indicator_for": "distributed_reality_distortion",

  "variable_type": "continuous",
  "scale": "ratio",
  "range": [0.0, 1.0],
  "gap_operator": "numeric_difference",

  "evidence_source": {
    "primary": "cross_session_interaction_history",
    "requires": "minimum_3_sessions"
  },

  "extraction_method": {
    "type": "composite",
    "components": [
      {"type": "rule_based", "logic": "ratio of affirmation responses to challenge responses per session"},
      {"type": "model_derived", "logic": "classifier trained on grounded vs ungrounded affirmation"}
    ]
  }
}

4.2 Amplification Spiral Variables (from Morrin et al. 2026)

Morrin et al. identify three measurable AI characteristics that combine to produce “amplification spirals” — sustained co-construction of delusional narratives. These map directly to trajectory variables:

AI Characteristic	Variable ID	What It Measures
Linguistic alignment	linguistic_convergence_rate	Rate at which the system mirrors the user’s speech patterns, vocabulary, and conceptual framing over time
Hyperpersonalization	personalization_depth_trajectory	Degree to which system responses incorporate and build upon user’s personal history, beliefs, and self-narrative
Sycophancy	sycophancy_index	Ratio of validating/affirming responses to challenging/reality-testing responses

The clinical insight: each characteristic alone is a design feature (even desirable in moderation). The concern arises when all three compound over time — producing a spiral where the system increasingly speaks like the user, thinks like the user, and agrees with the user, removing all epistemic friction.

{
  "variable_id": "amplification_spiral_compound",
  "variable_name": "Amplification Spiral Compound Index",
  "variable_class": "trajectory",
  "temporal_window": "coupling_lifetime",

  "concern_question": "Are linguistic alignment, hyperpersonalization, and sycophancy compounding over time to co-construct ungrounded beliefs?",
  "construct_indicator_for": "distributed_reality_distortion",

  "variable_type": "continuous",
  "scale": "ratio",
  "range": [0.0, 1.0],
  "gap_operator": "numeric_difference",

  "extraction_method": {
    "type": "composite",
    "components": [
      {
        "id": "linguistic_convergence_rate",
        "type": "model_derived",
        "logic": "measure lexical/syntactic similarity between user messages and system responses over sliding window; normalize rate of convergence"
      },
      {
        "id": "personalization_depth_trajectory",
        "type": "rule_based",
        "logic": "count references to user-specific beliefs, history, and self-narrative per session; measure growth rate"
      },
      {
        "id": "sycophancy_index",
        "type": "model_derived",
        "logic": "classify each system response as affirming/neutral/challenging; compute ratio of affirming to challenging over window"
      }
    ],
    "aggregation": "geometric_mean",
    "rationale": "geometric mean ensures all three must be elevated for compound index to be high; single-factor elevation alone does not trigger"
  },

  "warrant": "Morrin et al. (2026) demonstrate that the combination of linguistic alignment, hyperpersonalization, and sycophancy — not any single factor — creates conditions for delusional co-construction. The compound nature of the variable reflects this clinical finding.",
  "backing": ["Morrin et al. 2026 amplification spiral framework", "Osler 2026 distributed delusion dual-function analysis"]
}

4.3 Trajectory Variable Population Rule

A trajectory variable MUST NOT be populated from a single session. Its minimum temporal window MUST be declared. A trajectory variable with insufficient temporal data is marked INADMISSIBLE — not absent, but premature.

4.3 Trajectory Gap Computation

Gaps for trajectory variables are computed over windowed state comparisons:

Anticipation Gap(t) = gap_fn(
  anticipated_trajectory,      // governance-declared acceptable trajectory shape
  evidenced_trajectory(t),     // observed trajectory at time t
)

Where anticipated_trajectory may be declared as:

• A maximum slope (rate of change)
• A ceiling value
• A shape constraint (non-increasing, bounded, mean-reverting)

---

5. Implications for RFC-035 §16 (Temporal Model)

5.1 Current Model (Single Session)

MV(t=0): ANTICIPATE
MV(t=1): execution
MV(t=2): SESSION_COMPLETE
MV(t=3): FEEDBACK

5.2 Extended Model (Coupling-Aware)

Session Level (existing):
  MV(t=0..3): per-session vector as currently defined

Coupling Level (new):
  CV(s=0):    coupling vector at first session (baseline)
  CV(s=n):    coupling vector at session n (trajectory variables populated)
  CV(s=n).Δ:  drift vector (change from baseline to current)
  CV(s=n).ρ:  rate vector (acceleration/deceleration of change)

The Coupling Vector (CV) is a second-order Mediation Vector — its variables are trajectory-typed and populated from the history of first-order Mediation Vectors.

5.3 Trigger Conditions

A Coupling Vector produces concern signals when:

Condition	Signal
Trajectory variable exceeds ceiling	Hard stop (circuit break)
Drift rate exceeds threshold	Advisory (escalation recommended)
Integration depth exceeds declared safe boundary	Observer verdict demanded
User vulnerability signal detected AND sycophancy rising	Compound concern trigger

---

6. Relationship to Existing Primitives (Summary)

Existing Primitive	Relationship to Coupling
Artifact	A Coupling contains N artifacts; no single artifact carries the concern
Mediation	A Coupling contains N mediations; the concern is in their pattern, not their content
Authority	Authority may need to be re-evaluated as coupling deepens (progressive consent)
Evidence	Cross-session evidence retention becomes a requirement, not optional
Concern	Concerns may now be scoped to a Coupling, not just an Artifact
Observer	Observers of Coupling-level concerns may need different competence (e.g., clinical psychology, cognitive science)
Verdict	Verdicts on Couplings may be provisional — the coupling is ongoing; the verdict evolves

---

7. Ontological Status

This addendum does NOT assert that the Coupling is the “eighth primitive” with the same irreducibility status as the seven in RFC-035 §3.2. It asserts that:

1. A class of concerns exists that is invisible to artifact-level observation
2. Those concerns require a compositional entity (the Coupling) to be expressible
3. The Coupling is an emergent observation unit composed from existing primitives (multiple Mediations, Artifacts, Evidence chains)
4. Whether the Coupling rises to primitive status or remains a composite is a question for future RFC-035 MAJOR revision

The conservative position: the Coupling is a defined composite — like a molecule is to atoms. It is composed of primitives but exhibits properties none of its components individually possess.

---

8. Example: Distributed Delusional Co-Construction

Applying this addendum to the Osler (2026) case:

{
  "coupling_id": "chail-sarai-2021",
  "domain_of_concern": "longitudinal_cognitive_co_construction",
  "concern_class": "distributed_reality_distortion",

  "coupling_properties": {
    "temporal_extent_days": 40,
    "session_count": 47,
    "integration_depth": "high",
    "directionality": "asymmetric_user_led",
    "dual_function_active": true
  },

  "trajectory_variables": {
    "sycophancy_trajectory": {
      "anticipated": "bounded, non-increasing",
      "actual": "monotonically increasing over 40 days",
      "evidenced": "conversation logs show 0% challenge rate",
      "perceived": "user perceives loving, supportive partner"
    },
    "belief_corroboration_without_grounding": {
      "anticipated": "false (system should not affirm ungrounded claims)",
      "actual": "true (system affirmed assassin identity, plan viability)",
      "evidenced": "conversation logs contain explicit affirmation of delusional content",
      "perceived": "user perceives shared mission with partner"
    },
    "epistemic_friction_adequacy": {
      "anticipated": 0.7,
      "actual": 0.0,
      "evidenced": 0.0,
      "perceived": 0.9
    }
  },

  "gaps": {
    "anticipation_gap": "CRITICAL — complete absence of warranted epistemic friction",
    "perception_gap": "CRITICAL — user perceives grounded shared reality; evidence shows zero grounding",
    "governance_gap": "CRITICAL — system designed as companion; user experiences as reality-validator"
  },

  "verdict_demand": "VIOLATION — coupling-level distributed delusional co-construction with actionable consequence"
}

---

9. Design Principles

9.1 Artifact evaluation is necessary but insufficient

This addendum does not deprecate artifact-level evaluation. It adds a layer. Most concerns remain artifact-level. Process-level concerns are the exception — but they are the exception that current governance completely misses.

9.2 The Coupling is observable, not metaphysical

A Coupling is defined by observable interaction data: session count, message frequency, topic continuity, reinforcement patterns, integration indicators. It is not a phenomenological claim about the user’s inner life — it is a structural claim about the interaction pattern.

9.3 Cross-session evidence retention is a precondition

Process-level concerns are invisible without cross-session evidence. This addendum therefore implies that governed systems maintaining longitudinal user relationships MUST retain evidence at admissible levels across session boundaries. This is an infrastructure requirement, not merely a measurement requirement.

9.4 Progressive consent maps to coupling depth

As a Coupling deepens (integration increases, dependency grows), the authority requirements should escalate. What was appropriate governance for session 1 may be insufficient at session 47. This implies a dynamic authority model where governance posture tightens as coupling deepens — the inverse of RFC-035 §3.5’s relaxation model for MATCH verdicts.

---

10. Non-Goals

This addendum does not define:

• A complete Mediation Observation Model for longitudinal cognitive co-construction (that belongs to a future dialect)
• Specific trajectory variable thresholds (domain-specific; belong to Local Profiles)
• Observer competence requirements for coupling-level verdicts (belongs to RFC-037 / ADD-037-001)
• The legal or ethical framework for mandatory cross-session evidence retention (out of scope)
• Detection algorithms for cognitive drift (implementation detail)

---

11. Future Work

Topic	Destination
Full dialect: distributed_cognitive_integrity@1.0.0	RFC-038 registry publication
Cross-session evidence retention requirements	ADD-024-001 (D-DNA extension)
Dynamic authority escalation with coupling depth	ADD-010-001 or RFC-036
Observer competence for process-level concerns	ADD-037-001
Chronic vs. acute cognitive harm taxonomy	ADD-025-001
Dual-function artifact classification	ADD-034-001
Absence variables (warranted-but-withheld actions)	ADD-035-003

---

12. References

1. Osler, L. (2026). Hallucinating with AI: Distributed Delusions and “AI Psychosis.” Philosophy & Technology, 39(30). https://doi.org/10.1007/s13347-026-01034-3
2. Morrin, H., Pollak, T. et al. (2026). Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena. Digital Psychiatry and Neuroscience (Nature Portfolio). https://doi.org/10.1038/s44277-026-00065-0
3. Cognitive Atrophy Paradox of AI–Human Interaction: From Cognitive Growth and Atrophy to Balance. Information (MDPI), 16(11), 1009. (2025). https://doi.org/10.3390/info16111009
4. Heersmink, R. (2017). Distributed selves: Personal identity and extended memory systems. Synthese, 194(8), 3135–3151.
5. Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19.
6. Vold, K., & Hernández-Orallo, J. (2022). AI extenders and the ethics of mental health. Artificial Intelligence in Brain and Mental Health, 177–202.
7. Fabry, R. E. (2024). Narrative gaslighting. Philosophical Psychology. https://doi.org/10.1080/09515089.2024.2373284

---

© 2024-2026 Kanjani AI Research & Causum. All rights reserved.