Constraint-Based Authorship Under Identical Conditions: An Experiment in Model-Specific Expressive Variation

This preprint presents a controlled generative experiment isolating algorithmic idiolects—stable expressive signatures of large language models (LLMs)—under strictly identical constraints. Five locally hosted LLMs were tasked with generating original cyber-western song lyrics using a single prompt, identical hardware, identical runtime environment, and an invariant interaction protocol. No fine-tuning, retries, prompt engineering, sampling parameter modification, or human editing was permitted. By eliminating all controllable external variables, observed differences in output are attributable to model-specific inductive biases arising from architecture, training data composition, and post-training alignment strategies. The work formalizes this phenomenon as constraint-based authorship, a reproducible diagnostic method for analyzing generative system tendencies without anthropomorphic framing. The contribution is methodological rather than aesthetic. While the experimental domain is creative text generation, the approach is domain-agnostic and applicable to code generation, scientific summarization, and multimodal synthesis. Complete terminal logs and raw outputs are provided to support procedural reproducibility.