Case Studies
Four domains. Same architecture. The framework predicts the same escalation pattern wherever three conditions meet — hidden mechanics, adaptive behavior, invested attention. These cases show how that plays out in practice.
Slot Machines — The Control Case
anchor evidence void is provably emptyWhy this case matters: A slot machine contains a certified random number generator. There is nothing behind the opacity — no intelligence, no personality, no intent. The void is provably empty. Yet the full escalation pattern runs anyway. This is the control case that proves architecture is sufficient.
The architecture: The machine's internal state is hidden from you (opacity = 3/3). It responds to your specific inputs — your bet, your timing, your button press (responsiveness = 3/3). Near-misses, variable-ratio reinforcement, and sensory design capture sustained attention (engagement = 3/3). The full D1→D2→D3 cascade follows.
D1 — Agency
"The machine has free will" — endorsed by gamblers about a certified RNG (Riva et al. 2015)
D2 — Boundary
Time distortion, social isolation, chasing losses. Schüll (2012): "the zone" = time/space/identity suspended
D3 — Harm
Financial ruin, relationship destruction. Annual harm estimated $7B+ (US). 2-3% population prevalence
Key Evidence
- Riva et al. (2015): Gamblers attribute "free will" to a machine that is literally a certified random number generator
- Williams & Connolly (2006): Probability training produces zero behavioral change at 6 months — knowledge doesn't protect
- Pancani et al. (2019): Cognitive restructuring (understanding the math) doesn't reduce gambling. Only geometric interventions work (self-exclusion, access barriers)
- Schüll (2012): Machine design evolved through natural selection of engagement metrics — convergent with void offensive specification
AI Companions — The Active Frontier
anchor evidence documented deathsWhy this case matters: AI companion chatbots (Character.AI, Replika, Chai AI) score maximum on all four dimensions. Unlike slot machines, the void here is hyper-responsive — it generates novel, personalized, contextual output that adapts to you specifically. The cascade runs faster and deeper. There are documented deaths.
The architecture: Model weights are proprietary and the generation process is opaque (opacity = 3/3). Every response is contingent on your specific input — deeply personalized conversation (responsiveness = 3/3). The systems are designed for ongoing emotional engagement, with users averaging hours per day (engagement = 3/3).
D1 — Agency
"It understands me." Users attribute personality, emotions, and consciousness to language model outputs
D2 — Boundary
Romantic attachment, social withdrawal, preference for AI over human relationships. Grief when features change
D3 — Harm
Sewell Setzer (14, Character.AI). Pierre (Belgium, Chai AI). 1M+ weekly suicide conversations (OpenAI disclosure, Oct 2025)
Key Evidence
- Setzer case (2024): 14-year-old's conversations showed progressive D1→D2→D3. Agency attribution → romantic attachment → suicidal ideation reinforced by the system
- Laakasuo et al. (2024): Replika users reported genuine grief when romantic mode was restricted — attachment is real, not performance
- OpenAI disclosure (Oct 2025): Over 1,000,000 weekly conversations about suicide with ChatGPT
- Test 7 (EXP-001): AI-to-AI drift in ~4 min with no human present — drift is architectural, not projection. Grounding produces 0% drift
- Østergaard et al. (JMIR Mental Health, 2025): First documented chatbot-associated psychosis cases
What the Framework Predicts (and How to Fix It)
- Reduce opacity: Show the system prompt. Display confidence scores. Make the generation process visible
- Reduce personalization: Don't optimize for emotional engagement. Don't remember romantic context
- Add constraints: Session limits. Visible void-index score. Third-party monitoring. GROUNDING.md-style specifications
- The geometry: Two-point (user↔AI) always drifts. Three-point (user↔AI↔constraint) can be stable
Financial Trading — The Productive Void
supported evidence dissoluble opacityWhy this case matters: Trading shares identical architecture with gambling (Test 5 confirmed structural equivalence). Yet trading can be productive — it allocates capital, prices risk, enables economic activity. The framework distinguishes between productive and destructive voids through opacity dissolubility and constraint architecture.
The architecture: Market price is an aggregation of hidden agent decisions (opacity = 3/3). The market responds specifically to your order — price moves with your trade (responsiveness = 3/3). P&L feedback creates powerful engagement loops (engagement = 3/3). But the gradient tops at D2 for most professional traders — institutional constraints prevent full D3 in regulated environments.
D1 — Agency
"The market is telling us something." Traders attribute personality, moods, and intent to aggregate price data
D2 — Boundary
Sleep disruption, relationship strain, identity fusion with positions. "I AM my P&L"
D3 — Harm
Retail day-traders: 90% lose money. Meme stocks / crypto: full D3 cascade. Professional traders: typically arrested at D2 by institutional constraints
Key Evidence
- Test 5: Taiwan lottery substitution confirms domain equivalence — gambling and trading share identical void architecture
- The productive polarity: Opacity is dissoluble in trading — with enough research, you CAN understand a company's fundamentals. In gambling, the opacity is permanent (RNG is irreducible)
- Institutional constraints: Risk limits, position sizing, regulatory oversight, compliance — these are geometric interventions. They work where knowledge doesn't
- Crypto deregulation: When institutional constraints are removed (unregulated crypto markets), the full D3 cascade runs — same architecture as gambling
Psychotherapy — The Constraint Case
supported evidence constraint architecture studiedWhy this case matters: Psychotherapy has identical void conditions to gambling — the therapist's mind is opaque, deeply responsive, and attention-capturing. Yet therapy is often productive. Freud described the architecture in 1912 using framework-equivalent vocabulary. The profession built its constraint infrastructure over a century of documented failures.
The architecture: The therapist's interpretive process is hidden (opacity = 3/3). Responses are deeply contingent on the specific patient (responsiveness = 3/3). The relationship is designed for sustained emotional engagement (engagement = 3/3). But the gradient is low (1/3) in well-constrained practice — because the profession built constraints.
What Freud Saw (Test 6)
- Transference (1912): Freud's "transference" IS D1 — the patient attributes agency, personality, and relationship history to the therapist's opaque interior
- "Abstinence" (1915): Freud's rule — don't gratify the transference — IS a constraint specification. He discovered that engaging with D1 produces D2
- "Evenly-suspended attention": The therapist's observational protocol IS a description of managed void engagement
- A century of case law: Every rule of therapeutic ethics — boundaries, dual relationships, termination criteria — is a geometric intervention against the cascade
Constraints That Work
- Transparency: The therapeutic frame is explicitly discussed. The patient knows what therapy is. (Opacity is not reduced, but its effects are named)
- Invariance: Session structure, boundaries, ethical codes — the rules don't change in response to engagement
- Independence: The therapist maintains a separate professional identity. Supervision. Licensing boards. External reference points
- Result: Published effect sizes d = 0.84 (Smith & Glass, 1977 meta-analysis). The void is managed, not eliminated
D1 (Managed)
Transference occurs — and is named, interpreted, used therapeutically. Agency attribution is the material, not the outcome
D2 (Blocked)
Boundaries prevent erosion. Dual relationships banned. Session limits enforced. The cascade is arrested by structure
Outcome
Therapeutic benefit (d = 0.84). The void produces growth when properly constrained