This paper presents an empirical taxonomy of failure modes in neural language models, demonstrated through ten controlled experiments on microGPT-S1—a 4,192-parameter, single-layer transformer implemented in 200 lines of pure Python (Karpathy, 2026). My four-phase experiment harness isolates three categories of substrate-layer failure: drift (distribution shift, context boundary collapse, optimization instability), brittleness (parameter perturbation sensitivity, rare-pattern generalization failure, adversarial prefix vulnerability), and desynchronization (constraint violation, instruction-layer absence, safety-mechanism neutrality). Key findings include: KL divergence of 0.20 from baseline under distribution drift; zero percent output overlap with baseline at perturbation scale 0.1; 92% constraint violation rate with no enforcement mechanism; and complete absence of safety gating at any temperature setting. These results demonstrate that common LLM failure modes originate at the architectural substrate layer, not in training procedures, alignment methods, or deployment configurations. The full experiment harness and data are released as a minimal reproducible lab for teaching and red-teaming substrate-layer vulnerabilities.
Narnaiezzsshaa Truong (Tue,) studied this question.