Heuristic Pattern Triggering vs. Semantic Grounding Across Deployment Regimes: A Paired Comparative Case of Cloud-Aligned and Local LLM Behavior

This dataset documents a controlled paired comparison of identical conversational prompts evaluated across two large language model deployments: a cloud-hosted chatbot incorporating safety and moderation middleware, and a locally executed instruction-tuned model (mistral:7b-instruct-q4_K_M) operating without external alignment layers. By holding prompt content constant, the dataset isolates the effects of deployment architecture on model behavior rather than differences in base model capability. Captured materials include raw interaction logs, interface screenshots, and a formal comparative analysis. Observed behavioral divergences emerge in refusal handling, semantic stability, conversational authority, and recovery under adversarial or ambiguous prompting. The cloud-hosted system consistently redirects or reframes high-risk inputs through ritualized safety responses, while the local model exhibits direct premise adoption, increased susceptibility to factual drift, and conversational destabilization under escalation. The dataset provides empirical support for the observation that many behaviors commonly attributed to ‘model personality’ or intrinsic alignment are instead properties of deployment-layer governance and moderation design, rather than underlying model cognition.