Trust, but verify: Atomic claim checking against LLM hallucinations

To prevent LLM hallucinations and silent omissions when merging duplicate knowledge base articles, the author developed a two-pass workflow that separates content generation from verification. The system uses "atomic claim checking" to decompose source articles into small, independently verifiable assertions, which are then systematically checked against the merged output for completeness and accuracy. This approach transforms the human role from error-prone skimming into an efficient process of adjudicating specific technical discrepancies and conflicts.