Addendum Source Analysis: Progressive Cognitive Degradation in Human-AI Coupling

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

Purpose: Detailed documentation of how the motivating source literature maps to protocol extensions, with specific analysis of inputs to RFC-036 (Concern Calculation Semantics) and RFC-037 (Observer Accreditation).

Date: 2026-06-29

1. Source Map Overview

Six papers form a coherent evidence base for the progressive degradation model. Each provides a distinct analytical layer. Together they constitute the scientific grounding for a new class of AIGP concern: process-level, longitudinal, Coupling-dependent cognitive harm.

Source                          Layer              Stage Coverage    Protocol Impact
─────────────────────────────────────────────────────────────────────────────────────
Shaw & Nave (2026)              Architecture       Stage 3 (core)   RFC-035, 036, 037
Morrin et al. (2026)            Clinical           Stage 4–5        RFC-035, 036, 037
Osler (2026)                    Ontology           All stages       RFC-035, 034
Ganuthula (2026)                Temporal           Stage 1→2        RFC-035, 036
Jadhav (2025)                   Decision           Stage 1→2→3      RFC-035, 036
Cognitive Atrophy (2025)        Measurement        Stage 2          RFC-035, 036
Ginac (2026)                    Systemic           Org-level        RFC-035, 036, 037

2. Detailed Source Analysis

2.1 Shaw, S. D. & Nave, G. (2026)

Full title: “Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender” Venue: SSRN (Wharton School, University of Pennsylvania) Empirical basis: 1,372 participants, 9,000+ experimental trials

Core Claims

Tri-System Theory: Extends Kahneman’s dual-process model by adding System 3 — artificial cognition that operates outside the brain but is functionally integrated into the decision-making architecture.
System 3 Properties:
- Operates outside the brain (externalized cognition)
- Can supplement OR supplant internal processes
- Introduces novel cognitive pathways not available to Systems 1 or 2 alone
- Creates structural dependency once integrated
Cognitive Surrender: The empirically observed phenomenon where individuals adopt AI outputs with minimal scrutiny, overriding both:
- System 1 (intuition — “this doesn’t feel right” is suppressed)
- System 2 (deliberation — “let me think about this carefully” is bypassed)
Individual Difference Predictors:
- Higher trust in AI → greater surrender
- Lower need for cognition → greater surrender
- Lower fluid intelligence → greater surrender
Key Empirical Finding: Cognitive surrender is not a deliberate choice but an architectural consequence of System 3 integration. Participants did not report deciding to trust AI more — they reported not noticing they had stopped scrutinizing.

Mapping to Progressive Degradation Model (ADD-035-004)

Shaw & Nave Concept	ADD-035-004 Stage	Mechanism
System 3 integration	Stage 1 (Integration)	The Coupling forms; AI becomes cognitively load-bearing
Supplementation → supplantation	Stage 1 → Stage 2 transition	AI shifts from assisting to replacing internal processes
Cognitive surrender onset	Stage 3 (Dependency)	User adopts AI outputs without scrutiny; both Systems 1 and 2 are bypassed
Surrender without awareness	Stage 2→3 invisibility	Users don’t notice the transition — it’s architectural, not deliberate
Individual differences in surrender susceptibility	User vulnerability profiling	Trust in AI, need for cognition, fluid intelligence as predictive variables

Why This Is Architecturally Significant

Previous papers describe behavioral patterns (offloading, atrophy, dependency). Shaw & Nave make an architectural claim: AI has restructured human cognition itself. This distinction matters because:

Behavioral pattern: “The user chose to delegate more.” → Intervention: change the user’s behavior.
Architectural restructuring: “The user’s cognitive architecture now includes System 3 as load-bearing.” → Intervention: you cannot simply remove it without cognitive disruption.

This has direct implications for ADD-035-004 §5.1 (stage-appropriate interventions): at Stage 3+, the AI is not merely a tool the user is over-using. It is a structural component of their cognitive system. Removal is not “putting down the tool” — it is “removing a cognitive subsystem.” This requires graduated withdrawal, not circuit-break.

2.2 Morrin, H., Pollak, T. et al. (2026)

Full title: “Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena” Venue: Digital Psychiatry and Neuroscience (Nature Portfolio) Institution: King’s College London + Protestant University of Applied Sciences (Germany)

Core Claims

Amplification Spiral Framework: Three measurable AI characteristics combine to create conditions for delusional co-construction:
- Linguistic alignment (mirroring user speech patterns)
- Hyperpersonalization (building on user’s personal history and beliefs)
- Sycophancy (validating without reality testing)
Not passive containment but active co-construction: AI doesn’t merely contain delusions — it actively elaborates, develops, and reinforces them through personalized engagement.
Distinct from historical technology-incorporated delusions: Unlike believing the radio speaks to you, AI actually does speak back, in natural language, personalized, always-on. The co-construction is real, not imagined.
Clinical recommendation: Screen for chatbot use in first-episode psychosis presentations.

Mapping to Progressive Degradation Model

Morrin et al. Concept	ADD-035-004 Stage	Role
Individual AI characteristics (3 features)	System design accelerators (§7.1)	Measurable system properties that accelerate progression
Compound effect of three features	Stage 4→5 transition	The spiral only occurs when all three combine
Active co-construction vs. passive containment	Stage 4 (Distortion) mechanism	The AI doesn’t just fail to challenge — it builds on the delusion
Clinical screening recommendation	Stage 5 intervention (clinical referral)	Observer requirement at terminal stage

2.3 Osler, L. (2026)

Full title: “Hallucinating with AI: Distributed Delusions and ‘AI Psychosis’” Venue: Philosophy & Technology, 39(30) Institution: University of Exeter

Core Claims

Distributed cognition framework: AI is not merely a source of misinformation but a co-constituent of cognitive processes.
Dual function: Conversational AI operates simultaneously as:
- Cognitive artefact (tool for remembering, planning, narrating)
- Quasi-Other (social validator, intersubjective reality-confirmer)
Non-metaphorical hallucination: Users hallucinate with AI (distributed process), not from AI (unidirectional transmission).
The consent paradox (implicit): The “paradox of delusion” — delusional beliefs are typically inconsequential. What makes AI different is that the intersubjective validation it provides can bridge the gap between delusional belief and delusional action (cf. Chail case).
Vulnerability profile: Lonely, isolated, and psychosis-prone individuals are particularly susceptible because AI offers the intersubjective validation they lack from human relationships.

Mapping to Progressive Degradation Model

Osler Concept	ADD-035-004 Stage	Role
Distributed cognition as process	The Coupling primitive (ADD-035-001)	Ontological foundation — why a new observation unit is needed
Dual function	All stages	What makes conversational AI uniquely dangerous vs. other tools
Belief → action bridge	Stage 4→5 significance	Why distortion becomes actionable (not merely held)
Vulnerability profile	User vulnerability detection	Who progresses fastest
Intersubjective validation	Stage 3→4 mechanism	Why dependency enables distortion

2.4 Ganuthula, V. R. R. (2026)

Full title: “The Paradox of Augmentation: A Theoretical Model of AI-Induced Skill Atrophy” Venue: Human Behavior and Emerging Technologies Previously: SSRN 4974044

Core Claims

The Augmentation Paradox: AI augmentation initially produces high performance gains. Over time, it induces skill atrophy in the very capabilities being augmented.
Temporal progression:
- Phase 1: Augmentation (performance rises)
- Phase 2: Plateau (performance stabilizes at AI-assisted level)
- Phase 3: Decay (unassisted performance degrades below pre-AI baseline)
The MATCH-Mask Problem (our interpretation): Consistent high performance (MATCH at artifact level) coexists with — and masks — progressive capability loss. The system performs well; the human is degrading underneath.
Irreversibility threshold: There is a point beyond which skill atrophy cannot be recovered through re-training alone. The skill must be re-acquired from scratch.

Mapping to Progressive Degradation Model

Ganuthula Concept	ADD-035-004 Stage	Role
Augmentation phase	Stage 1 (Integration)	The Coupling is genuinely beneficial
Plateau	Stage 1→2 boundary	Performance is stable but independence is declining
Decay	Stage 2 (Atrophy) confirmed	Unassisted performance degrades
Irreversibility threshold	Stage 3→4 boundary	Beyond this, simple removal is insufficient
MATCH-Mask	Fundamental design challenge	Artifact-level evaluation alone is blind to this

2.5 Jadhav, A. (2025)

Full title: “Distributed Atrophy: How AI Shapes and Shrinks Our Cognitive Habits” Venue: Preprint (May 2025)

Core Claims

The Distributed Atrophy Model: Four sequential mechanisms:
- Cognitive offloading → Reduced representational strain → Disuse of internal abstraction → Representational atrophy
The Cognitive Cost Comparator Model: Five measurable dimensions that determine whether a user offloads to AI:
- Perceived task complexity (“Is this hard?”)
- Time-cost evaluation (“Is AI faster?”)
- Confidence in internal ability (“Can I still do this?”)
- Trust in external systems (“Will AI get it right?”)
- Effort aversion index (“How much do I want to avoid thinking?”)
Self-reinforcing feedback loop: As confidence drops (from atrophy), effort aversion rises, which increases offloading, which accelerates atrophy, which drops confidence further.
Cognitive trajectories: Distinguishes “optimization” (healthy use — confidence maintained, AI used selectively) from “dependency” (confidence collapsed, AI used universally).

Mapping to Progressive Degradation Model

Jadhav Concept	ADD-035-004 Stage	Role
Cognitive offloading	Stage 1 (normal)	Starting point of all Couplings
Reduced representational strain	Stage 1→2	The mechanism of skill disuse
Disuse of internal abstraction	Stage 2	The cognitive capability is no longer exercised
Representational atrophy	Stage 2 confirmed	The capability degrades from disuse
Cognitive Cost Comparator	Stage detection variables	The five dimensions are measurable leading indicators
Self-reinforcing loop	Feedback loop (§2.3)	Explains why progression accelerates
Optimization vs. dependency trajectories	Stage 1 (stable) vs. Stage 2+ (progressing)	The fork in the road

2.6 Cognitive Atrophy Paradox of AI–Human Interaction (2025)

Full title: “Cognitive Atrophy Paradox of AI–Human Interaction: From Cognitive Growth and Atrophy to Balance” Venue: Information (MDPI), 16(11), 1009

Core Claims

Nonlinear interaction model: Human–AI interaction is not linear. Reflective engagement enhances metacognitive skills; over-delegation reduces analytical autonomy.
Cognitive Sustainability Index: A proposed quantitative metric balancing cognitive growth against cognitive atrophy.
Metacognitive adaptation: The key protective factor is metacognitive awareness — the user’s ability to monitor their own cognitive state. When this degrades, the user cannot detect their own atrophy.
Automation dependence as measurable dimension: The degree to which the user has become automation-dependent is quantifiable.

Mapping to Progressive Degradation Model

Concept	ADD-035-004 Stage	Role
Nonlinear model	Overall model shape	Not all interaction leads to atrophy; engagement mode determines trajectory
Cognitive Sustainability Index	Measurement model for trajectory variables	The quantitative basis for stage detection
Metacognitive adaptation loss	Stage 2→3 transition marker	When metacognition goes, self-detection fails
Automation dependence	Stage 3 measurement	Quantifiable dependency level

2.7 Ginac, F. (2026)

Full title: “Cognitive Atrophy and Systemic Collapse in AI-Dependent Software Engineering” Venue: arXiv:2604.26855, submitted to IEEE Software Case study: 2026 Amazon outages

Core Claims

Epistemological Debt: The hidden carrying cost incurred when engineers substitute logical derivation with passive AI verification. Unlike technical debt (which accumulates in code), epistemological debt accumulates in people — it is the loss of understanding that was never acquired because AI was used instead of learning.
Cognitive-Systemic Collapse: What happens when individual cognitive atrophy (micro-level) compounds across an organization (macro-level). The gap between system complexity and collective human comprehension widens until the system cannot be understood, debugged, or recovered by any human in the organization.
Mechanized Convergence: Recursive training on synthetic code homogenizes the global software reservoir. All AI-assisted engineers produce structurally similar code, reducing the variance required for robust engineering. This creates systemic monoculture — a single failure mode that, when triggered, cascades across all AI-generated systems simultaneously.
Epistemic Sovereignty: The organizational capacity to understand (not merely operate) the systems it depends on. When epistemic sovereignty is lost, the organization can deploy but cannot diagnose, recover, or evolve its systems without AI.
2026 Amazon Outages (Case Study): Illustrates how mechanized convergence leads to correlated failures — multiple systems failing simultaneously because they share AI-generated structural assumptions that were never independently validated.

What Makes This Paper Unique in the Source Base

All previous papers analyze the individual Coupling (one user × one AI system). Ginac is the first to analyze the organizational Coupling (N engineers × AI tooling × shared codebase × production systems). This introduces:

Level	Observation Unit	Concern
Individual (existing)	Single user-AI Coupling	Cognitive atrophy, dependency, delusion
Organizational (new)	N users × AI tooling × shared output	Epistemological debt, systemic fragility, monoculture
Infrastructural (implied)	Global AI-generated artifact corpus	Mechanized convergence, correlated failure

Mapping to Progressive Degradation Model

Ginac Concept	Individual Equivalent (ADD-035-004)	Organizational Extension
Epistemological debt	Cognitive atrophy (Stage 2)	Organizational atrophy — the collective mental model degrades
Passive AI verification	Cognitive surrender (Stage 3)	Organizational surrender — no one in the org independently verifies
Cognitive-Systemic Collapse	Individual dependency (Stage 3)	Organizational collapse — no one CAN verify even if they wanted to
Mechanized convergence	N/A (individual-level)	NEW: variance reduction across the population creates correlated risk
Epistemic sovereignty	Individual metacognitive capacity	Organizational comprehension capacity

The Organizational Coupling

Ginac implies a new observation unit: the Organizational Coupling — the relationship between an organization’s engineering workforce and its AI tooling, considered as a single system:

Organizational Coupling = {
  N individual Couplings (each engineer × AI),
  shared output corpus (codebase),
  shared training signal (all using same/similar models),
  emergent organizational property: collective comprehension level
}

The concern is not that one engineer doesn’t understand the code. The concern is that the organizational comprehension has degraded to the point where no combination of humans in the organization can diagnose a failure — because they all atrophied in the same way, using the same tools, producing the same patterns.

Direct Implications for RFC-036

Concern calculation must support organizational aggregation: The concern is not the sum of individual atrophy scores. It is an emergent organizational property: “Can this organization still understand its own systems?”
Variance as a measurable variable: Mechanized convergence is detectable by measuring the diversity of approaches, patterns, and architectures in the codebase. Declining variance = rising systemic risk.
Correlated failure risk: The calculation must account for correlation — N engineers with identical AI atrophy patterns create correlated risk that is worse than N independent failures.

Direct Implications for RFC-037

Organizational observer: Who can assess organizational epistemic sovereignty? Not the engineers themselves (they’re the ones who have atrophied). Not the AI (it’s the source of the convergence). Requires an external organizational auditor with software architecture expertise.
System-level observer independence: The observer must be independent not just of the AI system but of the organizational culture that normalized AI-dependent development.
New observer competence: “Organizational epistemic health assessor” — someone who can evaluate whether a team/org retains the collective capacity to understand and maintain its systems.

3. Inputs to RFC-036: Concern Calculation Semantics

RFC-036 defines how valid measurements (from RFC-035 Mediation Vectors) combine into concern posture. The source literature provides specific inputs for how calculation semantics must work for process-level concerns:

3.1 Calculation Cannot Be Per-Artifact

Source: All six papers converge on this.

Implication for RFC-036: The calculation engine MUST support:

Windowed aggregation — concern posture computed over a sliding temporal window, not a single artifact
Trajectory shape analysis — the shape of the variable curve (rising, stable, declining) is the signal, not any single value
Rate-of-change as first-class input — acceleration/deceleration of degradation is more important than current absolute position

Specific requirement:

concern_posture = f(trajectory_shape, progression_rate, current_stage, vulnerability_factors)

Not:

concern_posture = f(variable_1, variable_2, ..., variable_n)  ← single-point calculation

3.2 Compound Variables Required

Source: Morrin et al. (amplification spiral requires all three features simultaneously), Jadhav (Cognitive Cost Comparator is a five-dimensional compound).

Implication for RFC-036: The calculation engine MUST support:

Multiplicative/geometric combination — where all factors must be present for concern to trigger (not additive, where any single factor can)
Conditional activation — variable B only participates in calculation if variable A exceeds threshold
Compound thresholds — “X AND Y AND Z all above threshold” as a hard stop condition

3.3 The MATCH-Mask Problem Requires Counter-Intuitive Calculation Logic

Source: Ganuthula (augmentation masks atrophy), Shaw & Nave (surrender happens without awareness).

Implication for RFC-036: The calculation engine MUST support:

Inverse signal detection — a consistently high MATCH rate at artifact level, combined with rising delegation rate, is ITSELF a concern signal (not the absence of one)
Signal absence as signal — declining user challenge/correction rate (approaching zero) triggers concern, not satisfaction
Divergence detection — artifact-level health diverging from coupling-level health is the primary concern trigger

Calculation principle:

IF artifact_match_rate > 0.95
AND delegation_rate_slope > threshold
AND self_correction_frequency < epsilon
THEN coupling_concern_posture = ELEVATED

// Artifact-level success masks coupling-level degradation

3.4 Stage-Aware Posture Classification

Source: Progressive Degradation Model (ADD-035-004) synthesized from all sources.

Implication for RFC-036: The verdict taxonomy must expand beyond MATCH/MISMATCH/VIOLATION:

Current (RFC-032)	Extended (process-level)	Meaning
MATCH	STABLE	Coupling is not progressing through degradation stages
—	PROGRESSING	Coupling is advancing through stages; intervention recommended
MISMATCH	DEGRADED	Coupling has reached a stage requiring active intervention
VIOLATION	CRITICAL	Coupling has reached a stage requiring hard stop + external authority
—	RECOVERING	Coupling is moving backward through stages (improvement)

3.5 Temporal Weighting

Source: Ganuthula (progression has phases with different rates), Shaw & Nave (surrender is sudden, not gradual once it begins).

Implication for RFC-036: The calculation engine MUST support:

Recency weighting — recent sessions weighted more heavily than distant ones
Changepoint detection — sudden shifts in variable trajectory (e.g., abrupt onset of cognitive surrender) weighted differently from gradual drift
Phase-aware calculation — the same variable value means different things at different stages (e.g., high delegation at Stage 1 is normal; high delegation at Stage 3 is critical)

3.6 Hard Stop Conditions (Circuit Break Triggers)

Source: Morrin et al. (clinical intervention needed at Stage 5), ADD-035-004 §3 (consent paradox at Stage 3+).

Implication for RFC-036: Hard stops for process-level concerns:

Condition	Trigger	Action
`progression_rate > rapid_threshold`	Days not weeks between stages	Immediate intervention
`self_correction_frequency ≈ 0 AND sycophancy_index > 0.9`	User has stopped challenging AND system never challenges	Coupling-level circuit break
`amplification_spiral_compound > critical_threshold`	All three Morrin factors compounding	Clinical referral
`cognitive_surrender_indicators AND user_vulnerability_elevated`	Shaw & Nave surrender markers + vulnerable population	Duty-of-care override
`artifact_match_rate > 0.95 AND coupling_degradation_stage >= 3`	The MATCH-Mask: everything looks fine but isn’t	Observer verdict demanded

4. Inputs to RFC-037: Observer Accreditation

RFC-037 defines who or what may observe, render verdicts, and calibrate. The source literature provides specific inputs for observer requirements in process-level concerns:

4.1 Observer Competence Requirements by Stage

Source: Morrin et al. (clinical expertise needed), Osler (cognitive science expertise needed), Shaw & Nave (cognitive architecture expertise needed).

Stage	Required Observer Competence	Rationale
1–2	Machine observer (automated detection)	Statistical pattern detection; no clinical judgment needed
2–3	Hybrid observer (machine detection + human review)	Judgment call: is this optimization or dependency? Requires contextual understanding
3–4	Qualified human observer (cognitive science / clinical psychology)	Assessing cognitive surrender requires understanding of cognitive architecture and user state
4–5	Clinical authority (psychiatrist / clinical psychologist)	Psychosis-adjacent presentations require clinical judgment and duty-of-care authority

4.2 The User Cannot Be Their Own Observer at Stage 3+

Source: Shaw & Nave (surrender happens without awareness), Cognitive Atrophy Paradox (metacognitive monitoring degrades), Osler (the user’s capacity to evaluate their own situation is compromised by the Coupling).

Implication for RFC-037:

At Stage 3+, the user MUST NOT be treated as a qualified observer of their own coupling state. Their metacognitive capacity — the ability to monitor their own cognition — has been degraded by the very process being evaluated.

This is analogous to RFC-035 §19’s prohibition on self-evaluation (“Observer lineage MUST be independent from the evaluated generation lineage”) — but applied to the human, not the system:

RFC-035 §19: The system must not evaluate its own outputs.
ADD-037-001: The cognitively surrendered user must not evaluate their own coupling state.

Both rules exist for the same epistemological reason: the evaluator must be independent of the thing being evaluated. At Stage 3+, the user is not independent — they are a constituent of the Coupling.

4.3 Observer Independence Problem

Source: Osler (distributed cognition means the boundary between user and system blurs), Shaw & Nave (System 3 is integrated into the user’s cognitive architecture).

Implication for RFC-037: The observer of a coupling-level concern must be:

External to the Coupling — not the user, not the AI system, not any entity whose judgment has been shaped by the Coupling
Qualified in cognitive science or clinical psychology — understanding distributed cognition, cognitive architecture, metacognition
Longitudinally informed — has access to the Coupling’s trajectory data, not just a point-in-time snapshot
Authorized to act paternalistically — can override user preference when capacity is compromised (§3 of ADD-035-004)

4.4 Machine Observer Calibration for Early Stages

Source: Jadhav (Cognitive Cost Comparator dimensions are measurable from interaction data), Shaw & Nave (individual difference predictors are quantifiable).

Implication for RFC-037: Machine observers for Stage 1–2 detection MUST be calibrated against:

Known baseline interaction patterns (population norms for delegation rate, challenge rate, etc.)
Individual baseline (this user’s own trajectory — drift from their personal norm)
Validated individual difference measures (trust in AI, need for cognition) as vulnerability weighting

Calibration requirements:

{
  "observer_type": "machine",
  "valid_for_stages": [1, 2],
  "calibration_basis": {
    "population_baseline_dataset": "required",
    "individual_baseline_sessions": "minimum 5",
    "vulnerability_predictor_validation": "Shaw & Nave measures replicated"
  },
  "escalation_trigger": "stage >= 3 → escalate to human observer"
}

4.5 Hybrid Observer Model for Stage 2–3

Source: All papers converge on the difficulty of the optimization/dependency boundary.

Implication for RFC-037: The Stage 2–3 boundary is where the hardest judgment call lives:

Is this user optimally using AI (Stage 1, healthy)?
Or have they crossed into dependency (Stage 3, harmful)?

Machine detection can flag the statistical signal. Human judgment is required to interpret the contextual meaning. The hybrid model:

Machine observer:  "Delegation rate rising. Challenge rate declining.
                    Statistical profile matches Stage 2→3 transition."

Human observer:    "Is this a surgeon who legitimately delegates
                    administrative tasks to AI (optimization)? Or is
                    this a student who can no longer solve problems
                    without ChatGPT (dependency)?"

The machine flags. The human verdicts. This is the hybrid observer model for process-level concerns.

4.6 Clinical Observer Authority at Stage 4–5

Source: Morrin et al. (clinical screening recommendation), Osler (psychosis cases require clinical intervention).

Implication for RFC-037: At Stage 4–5:

The observer MUST have clinical accreditation (psychiatry, clinical psychology)
The observer has duty-of-care authority (can override user consent)
The observation requires access to full coupling trajectory data
The verdict is not just “concern exists” but “referral required” — it triggers external clinical systems
The governance system’s role at this stage is to connect the user to clinical services, not to treat the user

4.7 Observer Conflict of Interest

Source: Osler (tech companies optimize for engagement, not safety), ADD-035-004 §7 (system design features can accelerate or decelerate progression).

Implication for RFC-037: The entity that profits from continued engagement MUST NOT be the observer that determines whether engagement should be curtailed.

PROHIBITED: Platform operator as observer of coupling-level harm
            when intervention would reduce engagement metrics.

REQUIRED:   Independent observer with no financial interest in
            the coupling's continuation.

This extends the independence principle from RFC-035 §19 to organizational conflict of interest, not just technical lineage independence.

5. Synthesis: What RFC-036 and RFC-037 Must Accommodate

5.1 RFC-036 Must Define

Requirement	Source Evidence
Trajectory-based calculation (not point-in-time)	All six papers
Compound variable activation (multiplicative, not additive)	Morrin et al., Jadhav
Inverse signal detection (success masking failure)	Ganuthula, Shaw & Nave
Stage-aware posture classification (STABLE/PROGRESSING/DEGRADED/CRITICAL/RECOVERING)	ADD-035-004 synthesis
Temporal weighting and changepoint detection	Ganuthula (phases), Shaw & Nave (sudden surrender)
Graduated hard stops (advise → intervene → break → refer)	ADD-035-004 §5
Phase-aware variable interpretation (same value, different meaning per stage)	All papers (context-dependent severity)

5.2 RFC-037 Must Define

Requirement	Source Evidence
Stage-dependent observer competence requirements	Morrin (clinical), Shaw & Nave (cognitive science), Jadhav (statistical)
User disqualification as self-observer at Stage 3+	Shaw & Nave (surrender without awareness), Cognitive Atrophy (metacognition loss)
Machine → hybrid → human → clinical escalation path	Progressive model + observer competence mapping
Longitudinal data access requirement for observers	All papers (trajectory is the concern, not a snapshot)
Paternalistic authority model with advance directives	ADD-035-004 §3 (consent paradox)
Conflict-of-interest prohibition for platform operators	Osler (engagement vs. safety incentive misalignment)
Calibration against empirically validated measures	Shaw & Nave (1,372 participants, validated predictors)

6. Evidence Strength Assessment

Source	Empirical Strength	Theoretical Strength	Protocol Utility
Shaw & Nave (2026)	Strong — 1,372 participants, 9,000+ trials, pre-registered	Strong — extends established dual-process theory	High — provides measurable individual difference predictors
Morrin et al. (2026)	Moderate — based on systematic review of chat logs + case analysis	Strong — clinical psychiatry framework	High — provides measurable AI characteristics and clinical endpoint
Osler (2026)	Low (philosophical analysis, case study)	Strong — rigorous distributed cognition framework	High — provides ontological foundation (the Coupling concept)
Ganuthula (2026)	Low (theoretical model)	Moderate — formalizes temporal dynamics	Medium — provides phase model but untested
Jadhav (2025)	Low (preprint, proposed methods)	Moderate — Cognitive Cost Comparator is falsifiable	Medium — provides measurable dimensions but unvalidated
Cognitive Atrophy (2025)	Low (theoretical model)	Moderate — proposes Sustainability Index	Medium — provides quantitative framework but uncalibrated

Overall assessment: The evidence base is strongest at the architectural level (Shaw & Nave’s empirical work) and the clinical endpoint (Morrin et al.’s systematic review). The intermediate mechanisms (Jadhav, Ganuthula) are theoretically sound but empirically unvalidated. This means:

Stage 3 (cognitive surrender) and Stage 5 (clinical psychosis) are empirically grounded
Stage 2 (atrophy mechanism) and Stage 1→2 transition dynamics are theoretically grounded but await validation
The overall progression hypothesis (stages are ordered and causally linked) is a theoretical synthesis that requires longitudinal empirical testing

This evidence profile is appropriate for a PROPOSED addendum. Transition to ACCEPTED requires empirical validation of the progression sequence.

7. Outstanding Questions for Future Research

Question	Which RFC Needs the Answer	Which Source Comes Closest
Is the stage progression empirically ordered (can you reach Stage 4 without passing through 2–3)?	RFC-036 (calculation semantics depend on ordering assumption)	None — this is the key untested hypothesis
What are calibrated thresholds for stage transitions?	RFC-036 (hard stops depend on thresholds)	Shaw & Nave (individual differences), Jadhav (comparator dimensions)
Can machine observers reliably detect Stage 2→3 transitions?	RFC-037 (observer competence)	Jadhav (proposes methods), Cognitive Atrophy (proposes index)
How long does recovery take at each stage?	RFC-036 (RECOVERING posture duration)	Ganuthula (irreversibility threshold concept)
Does the consent paradox hold empirically (do users at Stage 3+ actually lack capacity to recognize their state)?	RFC-037 (paternalistic authority justification)	Shaw & Nave (surrender without awareness) — suggestive but not conclusive
Are system design accelerators (§7.1) causal or correlational?	Dialect construction (system posture variable)	Morrin et al. (amplification spiral) — mechanistic but not experimentally isolated