CAR-RAG: Experimental Code, Data, and Reproducibility Archive
Description
This record contains archived experimental evidence associated with the CAR-RAG (Cost-Aware and Reliability-Driven Retrieval-Augmented Generation) research study.
The archive covers Experiments 22–26 and includes experimental notebooks, CSV results, summary statistics, confidence-interval outputs, a README, and a file-integrity manifest.
The experiments investigate cost-sensitive abstention, answer coverage, retrieval depth, and input-token efficiency in retrieval-augmented question answering using SQuAD 2.0 and FLAN-T5-base.
Results show that coverage-calibrated abstention can reduce cost-sensitive decision loss relative to native answering under the evaluated penalty assumptions. Adaptive retrieval can reduce input-token consumption relative to deeper fixed-depth baselines, but a statistically reliable decision-loss advantage over fixed k3 retrieval was not established.
This deposit provides archived experimental artifacts for inspection and reproducibility assessment. Independent end-to-end reproduction has not yet been demonstrated. The archive covers Experiments 22–26 only.