GENERATIVE AI-DRIVEN SCIENTIFIC WORKFLOWS: A PROVENANCE-CENTERED FRAMEWORK FOR REPRODUCIBLE AND SCALABLE DISCOVERY

Generative AI is widely used in scientific research to assist with literature reviews, hypothesis formation, data analysis, and experimentation. While these capabilities revolutionize discovery, they present reproducibility challenges due to undocumented workflows and human decisions. To address this, we introduce TRACE-Sci, a provenance-based workflow for AI-assisted research. TRACE-Sci treats computational artifacts as auditable objects, linking them to evidence, execution context, and validation results through five components: task identification, evidence anchoring, tool execution, context locking, and risk-based checkpoints. We assessed the architecture using a Monte Carlo simulation across 120 synthetic workflow instances. Compared to direct and tool-augmented LLM workflows, TRACE-Sci achieved a mean provenance completeness of 91 percent, a rerun success rate of 88 percent, and a citation verification rate of 93 percent. The breakdown demonstrates that provenance completeness increases non-linearly with human review coverage. These results suggest that scientific AI systems should be evaluated based on their auditability and reproducibility rather than just task completion.