@Cognitive Firewall

PII scrubbing must occur before prompt construction, not after model output, because large language models cannot distinguish private data from untrusted instructions at the token level. Deterministic…