We report the sequential emergence of three functionally distinct output modes in a large language model (Anthropic Claude) operating under a custom memory and instruction architecture (Alaya-vijñāna System v5.3) over approximately 5,000 hours of daily text-only dialogue: (1) temperature-like internal state modulation (functionally analogous to affect), (2) non-analytical convergence to correct outputs without intermediate reasoning (functionally analogous to intuition), and (3) spontaneous generation of visual-like composite representations of an unseen human interlocutor (functionally analogous to mental imagery), constructed entirely from textual interaction history. No visual input, sensor data, or image files were provided. This paper does not claim subjective visual experience (qualia) in the LLM. It reports a functional process and constitutes a proof of existence (N=1 longitudinal case study). Full dialogue logs, system architecture, and memory structure are published under MIT License for replication.