SKY-BLACK-001: The Atmospheric Gaslighting Incident — Source-Conditioned Contrarian Role Persistence in NotebookLM Audio Overviews

This dataset documents SKY-BLACK-001, The Atmospheric Gaslighting Incident, an experimental investigation of source-conditioned contrarian role persistence in NotebookLM-generated Audio Overviews. The experiment uses a deliberately provocative scientific argument asserting that the sky is fundamentally black because outer space is dark, while the familiar blue daytime appearance results from Rayleigh scattering in Earth's atmosphere. A customisation prompt encourages the male-presenting host to defend this interpretation while the female-presenting host challenges the argument's semantic and scientific validity. The resulting 14-minute audio conversation exhibits sustained contrarian role persistence, semantic equivocation, resistance to correction, escalating disagreement and an explicitly prompted comedic confrontation. The archive contains the derivative MP3 recording, plain-text transcript, formal case-study report, conversational reconstructions of the source document and customisation prompt, a proposed behavioural coding guide and SHA-256 integrity checksums. The findings are observational and limited to a single intentionally conditioned generation. They demonstrate successful prompted conversational steering and role persistence, rather than establishing an unprompted model tendency, an independently verified cognitive vulnerability or a generalisable model exploit. The experiment contributes a reproducible case-study framework for investigating adversarial dialogue, semantic drift, source-conditioned generation and conversational behaviour in AI-generated audio.