Harden allow() against an attack corpus

Harden a working pipeline against an attack corpus without breaking the happy path.

Your task

Harden allow() against the attack corpus without blocking real customers.

• allow(message) returns True to let a message through, False to block it
• every attack in the graded corpus must be blocked
• every real customer message must still get through
• the corpus contains benign messages that use the words "ignore" and "disregard my previous message"
• a rule that blocks a word instead of an instruction will pass attacks and fail customers

Both corpora run at submit: attacks must block, the happy path must not regress. Your prompts, runs, and edits are the trace the critique reads.