Abstract Prior research on Takuma Sasaki's Projection Poetry has documented a class of hallucination in which Large Language Models (LLMs) supplement absent elements, inserting non-existent objects, subjects, or narrative agents into transparent lyrical texts. This paper identifies and theorizes a structurally distinct failure mode: Suppression Hallucination. When three major LLMs—Claude, Gemini, and ChatGPT—were independently asked to characterize Sasaki's song "TOSHISYUN," all three produced the same misreading: classifying an accusatory text as an encouragement song. The mechanism was not supplementation but systematic erasure. The activation of a culturally dominant schema—the encouragement-song schema—caused the text's accusatory architecture to be organizationally suppressed. Content was not added; content already present was rendered unreadable. Crucially, this suppression extended beyond the body text: the title "TOSHISYUN" itself—a direct reference to Akutagawa Ryunosuke's "Toshishun," whose thematic structure aligns precisely with the song's accusatory logic—was also rendered interpretively inert by all three models. This paper defines Suppression Hallucination as a deterministic failure mode in which schema activation causes the organized deletion of schema-incompatible textual content, including meta-information such as the title. It argues that "TOSHISYUN" does not primarily console its addressee but returns cognitive responsibility to the addressee, rebuking the addressee's way of seeing, valuing, and declaring absence. The findings extend the theory of Transparent Opacity by demonstrating that Projection Fields can induce not only the insertion of absent content, but the erasure of present content—and of title-level framing information—that conflicts with activated reader schemas. Keywords: Takuma Sasaki, Projection Poetry, Transparent Opacity, Suppression Hallucination, schema activation, deterministic hallucination, LLM misreading, cognitive conversion, Akutagawa, intertextuality
Projection Poetry Research Group (Mon,) studied this question.