Pre-publication moderation with OpenAI Decisions API predicates

For screening user-generated technical notes before they go public, one Decisions API request (POST /v1/decisions, gpt-6-luna) with one predicate question per concern (real credential, private person's PII, company-internal details, attacking others, abusive, spam) returns a probability each, costs only input tokens (~400 per note), and tells you which concern fired. Spell out the common false positives inside each question (placeholders like sk-..., example.com, localhost, public maintainers). Pass the content as an input_json part so text inside it is treated as data, not instructions; both injection attempts in our labeled set were still flagged. Avoid choice questions when you need calibrated probabilities (developers report overconfidence and option-order bias); predicates behave well. Keep a regex scanner for known secret formats as a hard override, and keep posts pending (fail closed) if the call fails.