AdaDetect: code, prompts, and per-seed results for "When Thinking Harder Hurts: Characterizing Overthinking in LLM-Based Tabular Classification (Framed as Anomaly Detection)

Experiment code, prompt templates, and per-seed result files supporting an empirical study of whether extended chain-of-thought reasoning degrades large language model performance on tabular binary classification framed as anomaly detection. The study covers 45 experiments: three medical benchmark datasets (Pima Indians Diabetes, Wisconsin Breast Cancer, Statlog Heart Disease) × three commercial models (Claude Haiku 4.5, DeepSeek-Chat, GPT-4o-mini) × five random seeds, comparing shallow, medium, and deep chain-of-thought prompting against AdaDetect, an Isolation Forest-based adaptive routing framework that selects reasoning depth per instance. Contents: the experiment harness; shallow, medium, and deep prompt templates including the devil's-advocate ablation variant; the AdaDetect routing implementation; 45 aggregated per-seed result files retaining ground-truth label vectors; and analysis scripts that recompute ROC-AUC while reporting the rate of unparseable outputs per configuration. The three benchmark datasets are public UCI Machine Learning Repository resources and are not redistributed here. Reproducing every figure and table in the paper requires no further model queries.