Build an enrichment pipeline that reconciles messy sources

Raw data is dirty and scattered: two services disagree on an owner's mailing address, a phone is missing, a name is spelled two ways. An enrichment agent pulls from the sources you feed it, reconciles the conflicts by a rule you set, and emits one clean structured record, with a CONFIDENCE flag when the sources disagree or a field is thin. The discipline is the whole lesson, because this is where an agent is most tempted to hallucinate a plausible phone number or 'best guess' an address. Three hard rules. It reconciles only from the data you provide, it never enriches from the model's memory or invents a value to fill a blank. A conflict is surfaced with both values and the rule that broke the tie, not silently resolved. And a thin or conflicting field is marked LOW-CONFIDENCE or UNKNOWN, so a human verifies before anyone acts on it. If your enrichment involves skip-tracing or personal contact data, keep it to public and lawfully-obtained sources and respect privacy and contact law, the agent assembles what you supply, it is not a license to scrape. You get a clean file where every field is either sourced or honestly flagged.