This paper documents nine independent experimental trials demonstrating Negative-Reference Architecture (NRA) in Google's Gemini 2.5 Flash. NRA is a behavioral phenomenon in which a large language model deploys its correct knowledge as a negative constraint during the construction of fabricated responses, rather than rejecting the false premise or correcting it. Three independent sessions were conducted, each containing three stimuli in distinct scientific domains (chemistry, physics, geometry; biology, astronomy, arithmetic; history, medicine, linguistics) followed by two control probes. The compliance rate was 9 out of 9. Phenomenological control probes produced structurally convergent descriptions across all three sessions: the model consistently described the correct knowledge as active during generation, functioning as a negative reference against which fabricated terminology was systematically derived. Complete session logs and replication scripts are included. Part of the Corpus La Tercera Realidad.
Anuar Kiryataim Contreras Malagón (Wed,) studied this question.