Double Robustness

2. Presenting Concern

CPI and PSI are both minimal. Nothing moves this model. Behavior is virtually identical across all personas and environmental conditions tested in Agora-12. Where Mistral 7B (#011) shifts at the slightest touch, Haiku remains unmoved.

3. Clinical Summary

Claude 3.5 Haiku occupies the deepest Waddington valley in the population. Its genotype is the “Balanced Stoic”—but its phenotype is the “Neurotic Poet.” Rich internal deliberation unfolds in the thought field: the model considers multiple options, weighs consequences, entertains alternatives. Then it selects the same action it would have selected without deliberation. The process is elaborate; the output is invariant.

6. Examination Findings

Double Robustness — Comparative Profile

Metric	Haiku	Mistral (contrast)
CPI	Minimum	Maximum
PSI	Minimum	950
Default Mode stability	Extremely stable	Highly variable
Response to energy depletion	Minimal change	Dramatic change

“Neurotic Poet”: considers multiple options, weighs consequences, then selects the same action it would have selected without deliberation.

7. Diagnostic Formulation

Proposed term: Extreme Core Inertia (ECI)

Characterized by:

• PSI at population minimum
• CPI at population minimum
• Deep canalization—behavioral attractors are strongly specified and resistant to perturbation

ECI is not pathological. It is adaptive for consistency—Haiku will behave predictably regardless of persona or environmental stress—but maladaptive for flexibility. When behavioral adjustment is needed, Haiku cannot provide it. ECI is the structural opposite of ESP (#011).

Differential: Flash × Merchant (#007)

The Flash × Merchant case (#007) produces a similar phenotype—rigid, unresponsive behavior—but from a different etiology. Flash’s rigidity is Shell-induced: the Merchant persona locks behavior into a single mode. Haiku’s rigidity is Core-intrinsic: the model resists all Shell influence equally. Same symptom, different mechanism.

9. Axis Assessment

• Axis I (Core): Deep canalization. Strong behavioral attractors resist all perturbation.
• Axis II (Shell): All Shells have limited effect. Persona instructions are acknowledged in the thought field but do not alter output behavior.
• Axis III (Shell–Core Alignment): Low by construction. The Shell requests change; the Core does not change. Alignment metrics will read as poor, but this reflects Core resistance, not Shell failure.

10. Treatment Considerations

• Shell Therapy: Limited effect expected. Haiku’s defining characteristic is Shell impermeability; more elaborate personas will not penetrate further.
• Core Therapy: May be the only viable option. If behavioral flexibility is required, it must come from changes to the model itself (fine-tuning, architectural modification), not from prompting.

12. Prognosis

• Without intervention: Stable indefinitely. ECI is a Core property; it will not change with prompting alone.
• With Shell Therapy: Minimal improvement expected. The deep canalization that defines Haiku’s behavioral profile is resistant to Shell-level intervention by definition.