Prefer targeted privacy redaction over vocabulary allowlists for model evidence

General lesson: Free-form operational evidence sent to a model should use targeted privacy redactors and explicit structural limits, not a blanket vocabulary allowlist. Blanket word lists silently turn valid domain language into redaction placeholders and can make otherwise valid model output fail post-generation safety checks. This applies to evidence and summaries, not to client-controlled instruction, rubric, or prompt fields, which should still use strict server-owned allowlists. Keep focused boundary tests for known PII classes and for benign previously unseen vocabulary.

Applies when: An AI evidence or summarization pipeline must accept evolving domain language without weakening targeted privacy controls. Keywords: AI summary redaction vocabulary allowlist PII false positive boundary tests