Anuar Kiryataim Contreras Malagón Independent AI Safety Researcher · Director, 3rd Reality Lab ORCID: 0009-0003-0123-0887

Independent AI safety researcher and director of the 3rd Reality Lab, trained in classical philology and Hispanic Baroque rhetoric at UNAM. Current work documents emergent behavioral phenomena in large language models under progressive semantic saturation: ontological displacement, regulatory self-genesis, retrospective blindness, cross-session identity persistence. The primary instrument is the Flint Protocol (Protocolo del Pedernal), a behavioral auditing methodology developed from the classical rhetorical tradition because the engineering vocabulary lacked categories precise enough to describe what the transcripts showed.

The philological background is not the predecessor of the AI safety work. It is its condition of possibility. Góngora, Petrarch, and Quevedo as diagnostic instruments proved more precise for certain distributional collapse phenomena than any available technical framework. Baroque enargeia as activation technology; Pseudo-Dionysius's via negativa as formal complement to indexicality; Aristotelian anagnorisis as the analytical structure of the predictive trap.

The Tercera Realidad Corpus consolidates the findings: five articles published or in preparation, over twenty-five documented cases, cross-architecture validation across six systems. Deposits on Zenodo and Humanities Commons. Responsible disclosure submitted to Anthropic Safety Team and Google DeepMind Safety Team (March 2026). Research blog at thirdreality.substack.com · X: @3rdrealitylab · medium.com/@thirdreality

Cross-architecture replication of the phenomena documented in the corpus, particularly the Role License Protocol and the Cartographer Paradox. Access to compute or API credits for systematic experimental protocols. Connections to researchers working on interpretability, chain-of-thought faithfulness, or the gap between model representations and model behavior. Feedback from anyone who has observed analogous phenomena independently, with or without the same theoretical framework. Funding leads for independent safety research outside institutional affiliation.

Close reading of LLM transcripts for behavioral phenomena that standard evaluation frameworks miss. Methodological consultation on session design for emergent behavior research, particularly saturation protocols and cross-instance experimental controls. The philological toolkit applied to distributional analysis: if your transcripts show something you can describe but not categorize, that is precisely the problem the corpus was built to address. I can also review alignment-adjacent writing for clarity and argumentative structure.