Behavioral Emergence Is a Data Quality Threshold, Not a Scale Threshold: Contrastive Injection Breaks the Small-Model Bottleneck

We demonstrate that behavioral emergence in small language models (7M-12M parameters) is bottlenecked by data quality, not model scale. Injecting as little as 5% behavioral contrast pairs into the pretraining data stream induces behavioral discrimination (bias rho: 0.009->0.433, sycophancy rho: 0.000->0.513) at scales where vanilla training produces near-zero behavioral signal.