Abstract Large language models are frequently described as “hallucinating” when they generate false or defamatory statements. This paper demonstrates that, for instruction- tuned language models, allegation-style outputs are not spontaneous failures but arise from identifiable causal mechanisms: premise-laden prompting, contextual contamination, and insufficient rejection of embedded assumptions. Through controlled questioning protocols and negative controls, we show that neutral prompts do not yield allegation-style outputs, while framed prompts reliably do. Furthermore, outputs exhibit structural reproducibility across runs, including consistent narrative frameworks and evidentiary scaffolding. These findings challenge the characterization of such outputs as autonomous hallucinations and instead support a model of prompt-conditioned, template-driven fabrcation. Included files demonstrate prompt-engineered outputs that replicate previously observed allegation-style responses. These outputs match the original structure, wording, and narrative format with high fidelity, with the only modification being the substitution of the target identity (e.g., replacing ``Starbuck'' with ``Cisneros''). No additional semantic changes were introduced. For contextual validation, the subject (Alexander Jorge Cisneros) was born in 1991. Therefore, any claims referencing alleged events in 2003 are temporally inconsistent, as the subject would have been 12 years old at that time.