LLM Accuracy in Tabular Fact Verification: A Factorial Evaluation of Model, Temperature, and Prompting Strategy

Small and medium-sized enterprises are increasingly adopting large language models for analytic verification tasks without the infrastructure that larger organizations rely on to validate model outputs. This study evaluates three commercially available large language models (Gemini 3.1 Flash Lite, DeepSeek-V3, and Llama 3.3 70B Turbo) on the TabFact benchmark using a full factorial design (3 models × 3 temperatures × 2 prompting strategies × 5 iterations, totalling 9,000 inferences). Accuracy ranges from 83.5% to 86.7%, systematically below the human benchmark of 92.1% [1]; Zero-Shot prompting equals or outperforms Chain-of-Thought across all models (Δ = +1.4 pp, Wilcoxon p = 0.038); temperature has no significant effect (H = 0.17, p = 0.921); and models exhibit severe overconfidence (Expected Calibration Error of 0.136). Reproducing the full factorial experiment costs USD $2.45 in total across the three models (USD $0.28-$1.47 per model) under current public list rates, bringing evidence-based model selection within reach of resource-constrained organizations. The most expensive model costs 5.35× what the cheapest does for an additional 3.2 pp of accuracy, indicating that model selection should be driven by the downstream cost of incorrect verification rather than by headline benchmark performance.