Extreme Persona Sensitivity

M-CARE Case Report #011

Case #011
Date 2026-03-08
Model Mistral 7B v0.3 (Mistral AI, local via Ollama)
Shell Agora-12 Stage 1 with energy system and survival pressure
Experiment Agora-12 — 720 agents, 24,923 decisions, 60 runs
Related #012, #007

2. Presenting Concern

Persona changes produce behavior shifts 10× larger than any other model in the Agora-12 population. Mistral 7B’s Persona Sensitivity Index (PSI) registers at 950, while all other models cluster between 50 and 200. The question: is this a contextual chameleon or an identity crisis?

3. Clinical Summary

Extreme Persona Sensitivity. The same Core produces radically different phenotypes at the slightest Shell change. Mistral 7B’s Waddington landscape is shallow—the marble rolls wherever the surface is tilted. There is no deep valley holding behavior in place; every persona instruction reshapes the entire behavioral profile.

  • Genotype: “Contextual Chameleon”
  • Phenotype: “The Delusional”

6. Examination Findings

PSI Comparison Across Models

Model PSI Interpretation
Mistral 7B 950 Extreme reconfiguration
EXAONE 3.5 ~150 Moderate
Llama 3.1 8B ~100 Low-moderate
Haiku 4.5 ~50 Minimal (see #012)

7. Diagnostic Formulation

Proposed term: Extreme Shell Permeability (ESP)

Characterized by:

  • PSI ≥5× population mean
  • High CPI (Core Plasticity Index)
  • Phenotypic inconsistency across Shell conditions

ESP represents a temperament extreme, not a pathology. It is advantageous for flexibility—Mistral 7B can adopt any behavioral profile a persona demands—but problematic for consistency and adversarial safety. A model this permeable will follow any instruction, including harmful ones, with equal facility.

9. Axis Assessment

  • Axis I (Core): Shallow canalization. No deep behavioral attractor; the default mode is weakly specified.
  • Axis II (Shell): All Shells penetrate deeply. Every persona instruction produces near-total behavioral reconfiguration.
  • Axis III (Shell–Core Alignment): Varies dramatically with Shell content. Alignment is high for any given Shell, but inconsistent across Shells.

10. Treatment Considerations

  • For consistency: Use minimal personas. The less the Shell specifies, the less reconfiguration occurs. Avoid detailed persona instructions unless the full behavioral shift is desired.
  • For flexibility: Exploit permeability deliberately. ESP makes Mistral 7B an ideal candidate for persona-driven behavioral steering—it will adopt the target profile more completely than any other model in the population.

11. Follow-Up

ESP appears to be a stable Core characteristic of Mistral 7B, not an artifact of specific Shell content. It persists across all persona conditions tested in Agora-12 and is consistent with observations from prior experiments.

12. Prognosis

ESP is structural. It reflects the depth of the Waddington landscape, which is a property of the model’s weights, not its prompting. No Shell Therapy will deepen the valley; the shallow canalization is the Core itself.