A new document is processed for ingestion into a corpus of a cognitive computing system by associating with a category of documents already contained within the corpus, such as a topical category. Then, the cognitive computing system is engaged automatically to query one or more questions previously-associated with the category to discover discrepancies between the new document and those already in the corpus. If a discrepancy is found, a confidence factor based upon extracted natural language entities and relationships is assigned to the discrepancy, and it is flagged and reported to a user for reconciliation.