Two prior theoretical works in this series identified an unresolved tension in AI-assisted software engineering. Pereira (2026a) argued that naming conventions commonly adopted for human readability impose a hidden economic cost on LLM workflows through byte-pair encoding (BPE) tokenization, but offered only analytical projections. Pereira (2026b) proposed that structural thresholds derived from cognitive science—the Cognitive-Derived Coding Constraints (CDCC)—should also define an efficiency frontier for LLM processing, a convergence hypothesis left without empirical support. This paper closes both gaps.We frame LLM output generation as an economic production function:given code artifacts as inputs, the LLM produces output tokens subject to a capacity constraint. We conduct three controlled experiments using a reproducible Python pipeline. Experiment 1 measures token count differentials across naming conventions for a corpus of 200 enterprise event identifiers. Experiment 2 fits a log-log production function to 500 LLM responses across 100 Python functions stratified by cyclomatic complexity. Experiment 3 assesses whether efficiency rankings are robust across tokenizer vocabularies. Dot notation produces 1.12–1.20× more tokens than camelCase (𝑝 < 0.001), generating a projected cost differential of $54,499/year at enterprise API volumes. The production function yields an output elasticity of 𝛽 = 0.102 (𝑝 < 0.001), confirming strong diminishing marginal returns to complexity: a 1% increase in input tokens produces only a 0.10% increase in LLM output. Most critically, CDCC-compliant functions exhibit a 3.3× higher output/input ratio than violating functions (0.141 vs. 0.043, 𝑝 < 0.001), establishing CDCC thresholds as an empirical Pareto efficiency frontier. Efficiency rankings are perfectly consistent across all tokenizer pairs (Spearman 𝜌 = 1.000), confirming that camelCase’s advantage is universal. Together, the results demonstrate that structural choices governing code readability for human developers simultaneously govern LLM processing efficiency—a double dividend with direct implications for engineering practice. ∗Part of a three-paper series. Companion works: Pereira 2026a (Confirmation Bias in Post-LLMSoftware Architecture: Are We Optimizing for the Wrong Reader?); Pereira 2026b (CDCC: A Framework for Human–Machine Co-Design).