Your structured data extraction pipeline processes 1,200 insurance claim forms per day. Each extraction job runs autonomously and includes steps to validate parsed fields, write normalized records to a shared PostgreSQL staging table, and delete raw source files after successful ingestion. Three times in the past month, a misconfigured validation rule caused the pipeline to silently drop valid records — but the raw source files had already been deleted before the error was detected, making recovery impossible. The pipeline's system prompt currently instructs Claude to "process claims efficiently and clean up source files when done." What change would most effectively prevent irreversible data loss in future runs?