AI Text Pipeline: review the summary before normalizing it

Define the summary contract: target length, facts that must survive, allowed sources and how uncertainty is represented. Separate source text from instructions to reduce prompt-injection risk, remove unnecessary personal data, and cap input size. Preserve references from summary statements back to source spans so reviewers can verify them. Apply deterministic text normalization only after semantic generation; normalization must not erase negation or identifiers. Route low-confidence, truncated and policy-sensitive outputs to review instead of silently publishing them.