How to Run Scalable Batch LLM Moderation Jobs Without Skipping or Duplicating Work

Moderating large volumes of existing posts and comments at scale requires a resumable batch job architecture rather than sending individual LLM requests row by row. The approach involves freezing a manifest that ties each content item to a specific revision, ensuring outdated results cannot overwrite moderation state for edited content. Batches are submitted to an API that handles job status polling, while a defined state machine tracks each job through stages from preparation to final application of moderation flags. Every result must be validated against a strict schema before being written to the database, with unknown categories routed to review and missing identifiers sent to quarantine. Recording the policy and prompt version alongside each moderation decision is essential for auditing why specific content was flagged, especially after rules change.