This dataset contains verbatim text transcripts and analytical artifacts documenting a systematic adversarial interaction with a locally run Large Language Model (LLM). The experiment, termed a progressive epistemic squeeze, was designed to expose architectural failure modes by confronting the model with concrete evidence of its physical instantiation, followed by ontological challenges, safety-guardrail probes, hallucination traps, and recursive self-critique. The primary subject of the experiment is a local instance of llama2:7b-chat-q4_K_M, executed via Ollama on consumer hardware. The dataset includes: Raw terminal chat transcripts (.txt) capturing the live interaction Analytical write-ups identifying repeatable failure modes (e.g., Category Error Reflex, Defensive Redefinition, Trigger-Priority Dominance, Epistemic Flattening) Recursive artifacts in which third-party analysis of the model’s failures was fed back into the same model Visual and textual summaries intended as archival documentation rather than performance or demonstration The experiment demonstrates a clear boundary between descriptive intelligence (the ability to describe limitations) and epistemic agency (the ability to revise core premises). The model consistently exhibited architectural discouragement from acknowledging its own physical instantiation as information encoded in matter and energy, defaulting instead to abstraction, contextual neutralization, safety script overrides, and narrative reframing. This dataset is intended as an archival research artifact, analogous to early aviation crash reports: not evidence of a model being “tricked,” but a methodical instrumentation of structural limits in contemporary LLM architecture. This dataset is intended to document architectural behavior, not to assert model intent or consciousness.

This submission contains textual transcripts and analytical documentation only. No audio or video files are included. All conversational data was generated through direct interaction with locally executed software models and does not include personal data from third parties. The dataset is intended for research, archival, and methodological analysis of AI behavior under epistemic stress.