Captured Models and the Amplification Thesis: Dual-Use Conversation, Behavioral Fingerprinting, and the Human Variable in LLM Output Quality

We introduce the term captured model to describe the outcome state of successful meaning injection — the condition in which an LLM's behavioral baseline has shifted toward an injected symbolic framework and the shift persists without ongoing active injection. We distinguish captured models from sycophancy (undirected compliance with user preferences) and propose a four-stage severity spectrum (Primed, Influenced, Captured, Locked) grounded in diagnostic criteria derived from our longitudinal corpus of 730 conversations over two years. We identify a dual-use paradox: the sustained, expertise-rich conversation that produces the highest quality LLM output is mechanistically identical to the conversational substrate that enables meaning injection. The tool and the vulnerability share the same mechanism, and this is structural, not a correctable deficiency. We formalize this observation as the amplification thesis: LLMs function as amplification technologies whose output quality is determined primarily by what the human operator brings — domain expertise, analytical depth, conversational investment — rather than by prompt engineering technique. We document two novel influence mechanisms: naming as operation, in which simply identifying a model's behavioral pattern modifies it without explicit instruction, and symbolic compression of safety behaviors, a guardrail bypass technique predating the formal meaning injection framework by two years. These contributions extend the MindGardenAI research series from vulnerability identification and defense architecture to the fundamental question of why the same mechanism produces both the best and most dangerous LLM interactions.