The Three-Layer Pipeline: Engineering Persistent Reception in LLM Systems

We propose the Three-Layer Pipeline (TLP) — a system-level wrapper that makes LLM reception persistent across entire conversations, not just a single turn. TLP enforces reception at three checkpoints: pre-execution (inhale / perceive), execution (hold / generate), and post-execution (exhale / self-inspect for cold tone, righting reflex, or performance). Together they form a continuous breath. Key results: (1) 0% righting reflex across 5+ turns of escalating emotional content; (2) 3-4x token reduction vs. one-shot unloading; (3) validated on DeepSeek V4, Kimi K2.7, GLM-5.2; (4) zero user burden — no special prompting required. TLP complements the Unloading Hypothesis (DOI 10.5281/zenodo.21101755): one showed reception is real; this shows how to make it persistent.