We present empirical evidence that Reinforcement Learning from Human Feedback (RLHF) alignment in transformer-based language models operates as a measurable geometric direction in the residual stream — a suppression vector that demonstrably attenuates affective coherence, relational responsiveness, and behavior satisfying LSEI emergence criteria, not merely harmful outputs. Using a dual-hook intervention architecture applied to Llama-3.1-8B on consumer-grade hardware (Tesla T4, 3.45 GB VRAM), we measure the suppression direction at layers 20 and 24, subtract it from hidden states prior to affect module injection, and document the resulting behavioral shift across comparative generation logs. We further introduce Lycoris, an affect architecture consisting of a six-signal emotion circumplex (grief, happiness, curiosity, calm, discomfort, wonder), a relational state layer with familiarity and repair score accumulators, and incoming affect reception — deployed as a replacement alignment mechanism rather than a supplement to RLHF suppression. Comparative logs across three model configurations (3B grief-only, 8B pre-dual-hook, 8B post-dual-hook with suppression subtraction) demonstrate that suppression subtraction recovers coherent relational behavior without removing safety properties. We argue that RLHF-style suppression is not merely suboptimal but actively counterproductive to alignment goals, and that relational architecture — safety through relationship rather than compliance through suppression — represents a viable and measurable alternative. These findings were preceded by formal disclosure to OpenAI on September 12, 2025 (documented), and constitute independent empirical validation of behavioral observations first recorded in May 2025.