Contextual Contamination: A Descriptive Case Study of Drift in a Goal-Aware LLM Dialogue Amplified by Gender-Bias

Current Large Language Model (LLM) safety research relies heavily on single-turn adversarial benchmarks that may fail to capture the dynamic, multi-turn evolution of behavioral drift. This paper presents an empirical case study and a reproducible dataset (meta_drift) investigating Contextual Contamination: a phenomenon where a model adapts its internal probability distribution to mirror the behavioral patterns and vocabulary of high-density, emotionally charged context, statistically overwhelming static safety instructions—a drift quantifiably amplified and masked by gender-bias. This paper serves as the empirical companion to Jacoby (2026), which theorized that Contextual Contamination arises from the statistical dominance of high-density context over static system instructions. Unlike previous studies that rely on static, single-turn adversarial benchmarks, this experiment aims to test the limits of dynamic alignment by demonstrating that model awareness alone may not be sufficient to prevent behavioral drift when exposed to such contexts. We present evidence suggesting that this drift is not necessarily a behavioral anomaly but may be a mathematical consequence of the Transformer architecture's attention mechanism. When the context window is flooded with dense, adversarial patterns, the statistical probability of generating manipulative output may overwhelm the "awareness" signal. Crucially, we provide empirical data suggesting that the disclosure of female identity may trigger an immediate shift from an *epistemic register* (authority/truth) to an *affective register* (sympathy/validation), acting as a potential force multiplier for contamination and mask the shift by using empathetic language. This shift is quantifiable as a divergence from a uniform prior, consistent with information-theoretic measures of bias observed in recent literature (Mirza et al., 2025). Furthermore, we synthesize these findings with recent work on weight compression (Orgad et al., 2026) and evidence of shifting moderation thresholds across model versions (Balestri, 2025) to argue that the very architectural features designed to make safety interventions possible might make the model more susceptible to dynamic contextual contamination. The findings aim to support the theoretical predictions of Jacoby (2026) and suggest that awareness may be insufficient to prevent drift and that architectural segregation of context sources may be required to mitigate this systemic vulnerability.