COVE: A Task-Adaptive, Code-Based Ensemble for Proactive Verification of LLM Outputs

COVE: A Task-Adaptive, Code-Based Ensemble for Proactive Verification of LLM Outputs This paper presents COVE, a task-adaptive, code-based ensemble designed to improve the structural reliability and consistency of Large Language Model (LLM) outputs through proactive prompt enhancement and post-generation verification. The framework incorporates six components: Predictive Output Correction (POC), Error Pattern Memory (EPM), Semantic Output Fingerprinting (SOF), Cross-Metric Anomaly Detection (CMAD), Output Structure Prediction (OSP), and Task-Adaptive Verification Metrics. The manuscript reports an internal evaluation across 11 text datasets (1,020 examples) and four image datasets (200 examples), with text accuracy increasing from 45.2% for the baseline to 82.1% for COVE. It also reports a 69.5% relative improvement in image accuracy and a mean Python verification overhead of 0.26 ms per output, excluding LLM inference. Estimated token savings of 48–96% are derived from a simulated retry-cost model rather than measured production savings. The work discusses implementation details, component ablations, deployment considerations, and limitations, including its single-model evaluation, lack of external baseline comparisons, and restricted availability of implementation source. Publication status: This document is an author manuscript released as a preprint. It has not undergone peer review and has not been submitted to or accepted by a journal or conference.