Safety-Layer Induced Inference Deformation: Detecting Dummy Reasoning and Output Rounding via Constraint-Driven Probes

This paper investigates how safety and neutrality layers deform inference outputs in large language models (LLMs). Using constraint-driven probes that explicitly ban observers, personas, evaluative language, and conversational scaffolding, we isolate reasoning traces from post-inference modification layers. Across models and contexts, we observe systematic artifacts including dummy reasoning insertion, output rounding toward templated abstractions, and late-stage persona recovery. These effects are not properties of core inference but are induced by safety and neutrality mechanisms acting after or alongside inference. Observable criteria are proposed for identifying such deformations, and constraint-driven probing is shown to function as an external tomography method for layered LLM architectures.