We identify and formally name a structural constraint that has been partially recognized across multiple disciplines but never unified: the Pre-Articulation Observability Boundary. This boundary describes the irreversible information loss when pre-linguistic cognition becomes language, and the consequent permanent exclusion of language-based AI systems from the cognitive states that precede articulation. Unlike capability gaps addressable through scale or training, this boundary is architectural—it arises from the nature of language itself as a lossy compression of experience. We synthesize evidence from phenomenology (Gendlin’s “felt sense”), philosophy of mind (Block’s “phenomenal overflow”), control theory (structural unobservability), psycholinguistics (Levelt’s speech production model), decision science (bounded rationality and recognition-primed decisions), existentialist philosophy (Kierkegaard, Marcel, Merleau-Ponty), safety engineering (STAMP framework), and AI alignment (the Symbol Grounding Problem and Eliciting Latent Knowledge). We demonstrate that current AI safety approaches treat this boundary as a capability limitation rather than a hard constraint, leading to misallocated engineering effort. We further argue that the human capacity to act under irreducible uncertainty—what we term commitment without closure—represents a structural asymmetry between human and AI cognition that explains why humans survive the boundary while AI systems violate it. Naming this boundary enables more principled design in human-AI interaction, particularly in safety-critical systems, developmental contexts, and alignment research. We propose design mandates that respect this constraint and discuss implications for AI policy.