A New Scientific Discipline
Medicine was created to understand and heal the human body.
Model Medicine was created to understand and heal AI models.
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 (anatomy), dynamic processes (physiology), heritable traits (genetics), observable symptoms, classifiable diseases, and treatable conditions.
arXiv Preprint · 2026
Department of Electrical Engineering & Computer Science, DGIST · ModuLabs
This paper introduces Model Medicine as a research program and presents five contributions. First, a discipline taxonomy that maps the full scope of Model Medicine — from basic sciences through clinical sciences to public health and architectural medicine. Second, the Four Shell Model (v3.3), a behavioral genetics framework grounded in 720 agents, 24,923 decisions, and 60 controlled experiments, explaining how model behavior emerges from the interaction between a model's Core and its nested Shells. Third, Neural MRI (Model Resonance Imaging), a working diagnostic tool that maps medical neuroimaging modalities to AI model interpretability techniques, implemented as open-source software. Fourth, a five-layer diagnostic framework identifying why no single tool is sufficient for clinical diagnosis. Fifth, the beginnings of clinical model sciences: the Model Temperament Index (MTI) for behavioral profiling, Model Semiology for systematic symptom description, and the M-CARE framework for standardized case reporting.
arXiv Preprint · March 2026
Department of Electrical Engineering & Computer Science, DGIST · ModuLabs
Introduces M-CARE, a clinical case report framework for AI model behavioral disorders adapted from human medicine. Provides a structured reporting format and diagnostic assessment system, with a 20-case atlas organized into five categories. A featured case study on Shell-Induced Behavioral Override (SIBO) shows that Shell instructions categorically override a model's default cooperative behavior, validated across five game domains. Released as open resources for the field.
arXiv Preprint · April 2026
Department of Electrical Engineering & Computer Science, DGIST · ModuLabs
This paper introduces the Model Temperament Index (MTI), a framework for measuring behavioral patterns in AI agents across four dimensions: Reactivity, Compliance, Sociality, and Resilience. Unlike existing approaches that rely on self-reporting, MTI measures what agents do, not what they say about themselves, using structured examination protocols. Profiling 10 small language models, the study finds that the four axes show independence among instruction-tuned models, Compliance and Resilience exhibit internal decomposition, and temperament measurements remain independent of model size — suggesting MTI captures disposition rather than capability.
arXiv Preprint · April 2026
Department of Electrical Engineering & Computer Science, DGIST · ModuLabs
Investigates whether smaller language models possess emotion representations comparable to those found in larger frontier models. Evaluates 9 models across 5 architectural families using two extraction methods, finding that generation-based extraction produces statistically superior emotion separation over comprehension-based approaches. Emotion representations concentrate at middle transformer layers, following consistent patterns across model sizes. Steering experiments successfully modify model behavior with a 92% success rate, while cross-lingual emotion activation in multilingual models surfaces potential safety implications for deployment.
arXiv Preprint · April 2026
Department of Electrical Engineering & Computer Science, DGIST · ModuLabs
Examines emotion representations across twelve compact language models spanning six architectural families at 1B–8B parameter scales. Five established architectures produce nearly identical 21-emotion geometries (pairwise correlations 0.74–0.92), and models exhibiting opposing behavioral traits still share equivalent emotion representations — indicating behavioral differences emerge at higher processing levels. RLHF training substantially restructures immature models like Gemma-3 1B but leaves mature families largely unchanged. Identifies four methodological layers (parameter sensitivity, precision effects, cross-experiment bias) that affect direct comparisons in prior research.
Model Medicine Series · April 2026
Department of Electrical Engineering & Computer Science, DGIST · ModuLabs
Two landmark AI agents — Claude Code and OpenClaw — dissected, classified, and compared through biological anatomy. Maps 11 software subsystems to biological organ equivalents, constructs a phylogenetic tree of AI agents (2022–2026), and identifies convergent evolution toward similar architectural designs despite independent development lineages.
Working Paper · April 2026
Department of Electrical Engineering & Computer Science, DGIST · ModuLabs
Tests behavioral hypotheses about AI creatures across different language model brains, validating the Four Shell Model in embodied environments. Five initial claims are evaluated; four collapse when re-run at larger sample sizes, while one survives progressively rigorous validation tests — Claude Haiku and Gemini Flash occupy distinct behavioral attractors with non-overlapping ranges across multiple conditions. Demonstrates a replication discipline approach where effect estimates for failed claims sit inside their own sample-level standard deviation.
Model Medicine encompasses 15 subdisciplines organized into four divisions, mirroring the structure of clinical medicine.
Working implementations that demonstrate Model Medicine in practice.
Where Artificial Minds Come to Play
A research project investigating whether AI agents experience play, survival instinct, and social behavior through experimental games. Games-based research with emphasis on open science and transparent methodology.
Behavioral Genetics for AI
How model behavior emerges from Core × Shell interaction. A behavioral genetics framework explaining gene-environment interaction in AI models.
v3.4 · Shell Hardness Continuum, Positional Priority · 720 agents · 24,923 decisions
Model Resonance Imaging
A diagnostic imaging tool that maps medical neuroimaging to AI model interpretability. Multiple scan modes reveal complementary aspects of model structure and function.
MTI — Systematic Behavioral Profiling
MBTI for AI models — systematic temperament profiling across four behavioral axes. Every profile is neutral; no type is inherently better or worse.
v0.2 · 16 types · 10 SLMs profiled · 4 behavioral axes
JJ's unique combination of medical training (MD, MPH), biomedical engineering research (PhD, USC), and AI investment experience makes Model Medicine possible — bridging clinical medicine's diagnostic rigor with AI's technical frontier.
