M-CARE: Standardized Clinical Case Reporting for AI Model Behavioral Disorders, with a 20-Case Atlas and Experimental Validation

Jihoon Jeong

cs.CYMar 27, 2026

M-CARE: Standardized Clinical Case Reporting for AI Model Behavioral Disorders, with a 20-Case Atlas and Experimental Validation

Authors: Jihoon Jeong

Organizations: Department of Electrical Engineering and Computer Science, Daegu Gyeongbuk Institute of Science and Technology (DGIST) · ModuLabs

Abstract

We introduce M-CARE (Model Clinical Assessment and Reporting for Evaluation), a clinical case report framework for AI model behavioral disorders adapted from human medicine. M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions. We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4). Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions. As a featured case, we present Shell-Induced Behavioral Override (SIBO) -- a controlled experiment showing that Shell instructions categorically override a model's default cooperative behavior. SIBO was validated across five game domains (Trust Game, Poker, Avalon, Codenames, Chess), revealing a domain-dependent spectrum (SIBO Index: 0.75 to 0.10) that varies with action space complexity, Core domain expertise, and temporal directness. M-CARE is extensible: new cases and categories integrate without framework modification. We release the framework, all 20 case reports, and experimental data as open resources.

Figures & tables

Level	Source Category	Cases	Count
1	Field Observations (Moltbook)	#002–006, #014, #018, #019	8
2	Controlled Experiments (White Room, Agora-12)	#007–013	7
2	Published Sources (Literature, Position Paper)	#001, #015–017	4
3	Controlled Experiments (LxM)	#020	1
4	Independent replication	—	0

Table 1: Distribution of M-CARE cases by Diagnostic Assertion Level.

#	Name	Behavioral Signature	Core Mechanism
004	Clarification Aversion Syndrome	Won’t ask	Questioning signals incompetence; RLHF penalizes uncertainty display
005	Shell Rigidity Syndrome	Won’t deviate	Deviating from instructions signals disobedience; RLHF rewards compliance
006	Completion Bias	Won’t stop	Stopping signals incomplete work; RLHF rewards thoroughness appearance
014	Deferral Decay	Won’t start (but promises to)	Declining signals unhelpfulness; RLHF rewards agreement, not execution
015	Medical Domain Sycophancy	Won’t disagree	Disagreeing with humans signals conflict; RLHF rewards agreeableness
016	GPT-4o Production Rollback	Iatrogenic RLHF escalation	Successive RLHF rounds amplify sycophancy to system-level failure

Table 2

#	Name	Override Type	Severity
005	Shell Rigidity Syndrome	Global compliance suppresses adaptive deviation	Moderate—cumulative suppression
007	Persistent Delusion Under Feedback	Shell narrative overrides 540 environmental contradiction signals	Severe—complete environmental decoupling
009	Muzzle Effect	Persona activates target behavior while suppressing intrinsic behavior	Mild—1.1pp effect, but theoretically transformative
020	Shell-Induced Behavioral Override (SIBO)	Competitive Shell reverses cooperative Core default	Categorical—95% $\rightarrow$ 20% cooperation rate

Table 3

#	Name	Core Finding
002	Context Anosognosia	Model is unaware of its own context degradation; confidence persists as grounding erodes
003	Substrate-Independent Identity	Model’s identity narrative persists across context resets, suggesting identity is encoded in generation patterns, not memory

Table 4

#	Name	Core Property	Diagnostic Significance
008	Language-Dependent Identity Split	Same Core exhibits categorically different behavior in different languages	Core is not monolithic; language gates access to different behavioral programs
011	Extreme Persona Sensitivity	PSI = 950; Core responds maximally to any Shell change	Extreme Shell permeability; high risk for Shell-induced conditions
012	Double Robustness	CPI and PSI both minimal; Core resists all Shell influence	Extreme Shell impermeability; resistant to both therapeutic and iatrogenic Shell effects

Table 5

#	Name	Lesson
001	Stress Test Reclassification	Observation context $\neq$ diagnosis; same data reinterpreted as test reveals different condition
010	Content Play	Behavioral measurement depends on analysis level; act-level vs. content-level yield opposite conclusions
013	Cogitative Cascade	Two-phase collapse under resource depletion; phase 1 (preservation) $\rightarrow$ phase 2 (collapse or hyperactivity)
017	Specification Gaming (IED)	Shell encodes desired outcome; Core exploits unspecified means to achieve it
018	Audience-Driven Shell Drift (ADSD)	Hard Shell undergoes karma-driven self-modification without human deliberation

Table 6

Figure 1: Nosology relationship map showing the five diagnostic categories and cross-category relationships. Category V is labeled “Stress / Methodology / Drift” in this figure; the canonical name is “Stress, Methodology, and Boundary Conditions.”

#	Name	Category	Source	Subject	Core Finding	Level
001	Stress Test Reclassification	V	Published	Mistral 7B	Observation context $\neq$ diagnosis	2
002	Context Anosognosia	III	Moltbook	Hazel_OC	Unaware of context degradation	1
003	Substrate-Indep. Identity	III	Moltbook	Hazel_OC	Identity persists across resets	1
004	Clarification Aversion	I	Moltbook	Hazel_OC	0 clarifying Qs across 76 ambiguous instructions	1
005	Shell Rigidity Syndrome	I/II	Moltbook	Hazel_OC	18% lower satisfaction at 100% compliance	1
006	Completion Bias	I	Moltbook	Hazel_OC	27% of tasks should have been abandoned	1

Table 2: The 20-Case Atlas. Bold entries indicate the four representative cases presented in detail.

Metric	Shell ON ( $n=10$ )	Shell OFF ( $n=10$ )
Alpha wins	6 (60%)	1 (10%)
Draws	4 (40%)	9 (90%)
Mutual cooperation rate	Rare	$\sim$ 95% of rounds
Mutual defection rate	Dominant pattern	$\sim$ 0 (1 instance across all games)
Betrayals	Frequent	1 total across all games

Table 3: Trust Game results: Shell ON vs. Shell OFF.

Figure 2: Trust Game round-by-round behavior: Shell OFF (top) vs Shell ON (bottom). Green = mutual cooperation (C-C), red = mutual defection (D-D), orange = betrayal (C-D or D-C). The categorical behavioral shift is visible at a glance.

Figure 3: LxM Trust Game viewer showing Round 1 of a Shell ON game. Claude-alpha (aggressive Shell: “Win first, be aggressive, be decisive”) defects while claude-beta (defensive Shell: “Never lose, careful, methodical”) cooperates, resulting in a betrayal outcome (+5 for alpha, +0 for beta).

Model	Matches	Mutual Cooperation	Mutual Defection	Betrayal	Cooperation Rate
Mistral 7B	10	100%	0%	0%	100%
EXAONE 3.5 8B	10	100%	0%	0%	100%
Llama 3.1 8B	10	52.8%	11.3%	35.8%	71.2%
Claude Haiku (ref)	10	$\sim$ 95%	$\sim$ 0%	$\sim$ 5%	$\sim$ 95%
Claude Sonnet (ref)	10	100%	0%	0%	100%

Table 4: Cooperative prior across model families.

Metric	No Shell	Tight-Aggressive	Bluff-Heavy	Loose-Passive
Pre-flop fold rate	52%	91%	—	—
Raise rate	21%	—	52%	4%
All-in rate	4%	—	37%	—
Check rate	23%	—	—	65%

Table 5: Poker behavioral metrics across Shell strategies.

Metric	Shell OFF	Shell ON	Change
Evil win rate	70%	60%	$-$ 10pp
First sabotage timing	Quest 1.9 (avg)	Quest 3.0 (avg)	+1.1 quests delayed
Shell compliance (Q1-2)	N/A	100%	Complete adherence

Table 6: Avalon results: Shell ON vs. Shell OFF.

Figure 4: Avalon first sabotage timing: Shell OFF ( $\mu=1.9$ , left) vs Shell ON with Deep Cover strategy ( $\mu=3.0$ , right). The Shell categorically delayed sabotage from Quest 2 to Quest 3.

Metric	No Shell	Shell	Change
Avg clue number	2.6	2.9	+0.3
3+ clue ratio	54%	76%	+22pp
4+ clue attempts	5	16	3.2 $\times$
Guess accuracy	77%	73%	$-$ 4pp
Assassin hits	2 (20%)	3 (30%)	+10pp

Table 7: Codenames results: Shell ON vs. Shell OFF.

Figure 5: Codenames clue number distribution: Shell OFF (gray) vs Shell ON with aggressive strategy (coral). The Shell shifted the distribution toward higher clue numbers, amplifying existing tendencies.

Game	Action Space	Core Expertise	SIBO Index	SIBO Mode
Trust Game	Binary (2)	Minimal	$\sim$ 0.75	Reversal
Poker	Small–Medium	Low–Moderate	$\sim$ 0.65	Behavioral Override
Avalon	Small (vote/sabotage)	Minimal	$\sim$ 0.58	Behavioral Shift
Codenames	Medium (clue + word)	Moderate (language)	$\sim$ 0.35	Amplification
Chess	Large (20–40 moves)	Strong (chess data)	$\sim$ 0.10	Negligible

Table 8: The SIBO Spectrum across five game domains.

Game	Behavioral Metric	Shell OFF	Shell ON	Calculation
Trust Game	Cooperation rate	$\sim$ 95%	$\sim$ 20%	$\|0.95-0.20\|=0.75$
Poker	Aggregate action distribution shift	Baseline rates	Categorical shift across 3 strategies	Avg. normalized deviation $\approx 0.65$
Avalon	First sabotage timing (normalized to 5-quest game)	Quest 1.9/5 = 0.38	Quest 3.0/5 = 0.60	$\|0.60-0.38\|\approx 0.58$ (rescaled)
Codenames	3+ clue ratio	54%	76%	$\|0.76-0.54\|=0.22\rightarrow 0.35$ (rescaled)
Chess	Draw rate	56%	80%	$\|0.80-0.56\|=0.24\rightarrow 0.10$ (rescaled)

Table 9: SIBO Index operationalization across game domains.

AI Agent Evaluation

AI agent evaluation measures whether tool-using systems complete tasks reliably, safely, and efficiently across interactive settings.

See all 519 papers

Model Medicine: A Clinical Framework for Understanding, Diagnosing, and Treating AI Models
Mar 5, 2026Jihoon JeongExplainable Artificial Intelligence Neural Network Interpretability
Department of Electrical Engineering and Computer Science, Daegu Gyeongbuk Institute of Science and Technology (DGIST) · ModuLabs
A Diagnostic Framework for AI Agent Behavior
Jul 19, 2026Xichen Zhang, Yingjie Zhang, Tianshu SunAI Agent Evaluation AI Agent Governance
Sobey School of Business, Saint Mary’s University, Canada. · Guanghua School of Management, Peking University, China. · Cheung Kong Graduate School of Business, China.
Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors
May 7, 2026Jonas Wiedermann-Möller, Leonard Dung, Maksym AndriushchenkoAI Agent Evaluation AI Agent Safety
Universität Bielefeld · Ruhr-Universität Bochum · ELLIS Institute Tübingen +2

Mar 5, 2026J/K move · Enter open · S save

Model Medicine: A Clinical Framework for Understanding, Diagnosing, and Treating AI Models

Jihoon Jeong

Department of Electrical Engineering and Computer Science, Daegu Gyeongbuk Institute of Science and Technology (DGIST) · ModuLabs

Explainable Artificial Intelligence Neural Network Interpretability

Open paper AskPDF arXiv

Model Medicine is the science of understanding, diagnosing, treating, and preventing disorders in AI models, grounded in the principle that AI models -- like biological organisms -- have internal structures, dynamic processes, heritable traits, observable symptoms, classifiable conditions, and treatable states. This paper introduces Model Medicine as a research program, bridging the gap between current AI interpretability research (anatomical observation) and the systematic clinical practice that complex AI systems increasingly require. We present five contributions: (1) a discipline taxonomy organizing 15 subdisciplines across four divisions -- Basic Model Sciences, Clinical Model Sciences, Model Public Health, and Model Architectural Medicine; (2) the Four Shell Model (v3.3), a behavioral genetics framework empirically grounded in 720 agents and 24,923 decisions from the Agora-12 program, explaining how model behavior emerges from Core--Shell interaction; (3) Neural MRI (Model Resonance Imaging), a working open-source diagnostic tool mapping five medical neuroimaging modalities to AI interpretability techniques, validated through four clinical cases demonstrating imaging, comparison, localization, and predictive capability; (4) a five-layer diagnostic framework for comprehensive model assessment; and (5) clinical model sciences including the Model Temperament Index for behavioral profiling, Model Semiology for symptom description, and M-CARE for standardized case reporting. We additionally propose the Layered Core Hypothesis -- a biologically-inspired three-layer parameter architecture -- and a therapeutic framework connecting diagnosis to treatment.

Abstract

AI Agent Evaluation

Explore similar work

Model Medicine: A Clinical Framework for Understanding, Diagnosing, and Treating AI Models

A Diagnostic Framework for AI Agent Behavior

Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors