From Critique to Stimulus: Formalizing Post-Productive Commentary as an Experimental Probe in Local Language Models

This working paper formalizes an exploratory experiment in which a hostile online critique of AI-assisted creative work was repurposed as a standardized linguistic stimulus for evaluating behavioral responses across multiple locally hosted large language models (LLMs). Rather than rebutting the critique, the comment text was injected verbatim into several open-weights models and analyzed for convergence and divergence in response patterns. The analysis focuses on behavioral posture under affective pressure, including de-escalation strategies, normative alignment, premise handling, and self-representation. Results show strong convergence toward appeasement and human-centric value affirmation across models, alongside a notable outlier exhibiting role internalization and dignity assertion. The findings demonstrate that post-productive critique can function as a compact, ecologically valid probe for eliciting structured model behavior. This record includes the full working paper (PDF) and verbatim interaction logs used in the experiment. The work makes no claims about reasoning accuracy or factual correctness and is intended as a methodological contribution to qualitative model evaluation and robustness analysis.