From a 4-persona fresh-eyes UX review. Highest-leverage change — all four reviewers converged on it.
Problem
The headline (README line 8) promises prompt-injection defense. The honest admission that the default engine is a three-keyword substring filter lives in Limitations at line 113 — ~100 lines below the claim. A skeptic reads that gap as a walk-back:
"honesty that arrives after the claim reads as a walk-back; honesty that arrives before it is credibility."
This is the biggest credibility risk in a live demo: it invites the "this is a profanity filter with three entries" takedown before the reader ever reaches the caveat.
Fix
Insert one sentence immediately after The Solution (after line 20):
The interception point and the policy protocol are the product; the built-in SensitiveDataFilter is a demo default that matches three literal keywords — you write the policies for your own threat model.
Done when
From a 4-persona fresh-eyes UX review. Highest-leverage change — all four reviewers converged on it.
Problem
The headline (README line 8) promises prompt-injection defense. The honest admission that the default engine is a three-keyword substring filter lives in Limitations at line 113 — ~100 lines below the claim. A skeptic reads that gap as a walk-back:
This is the biggest credibility risk in a live demo: it invites the "this is a profanity filter with three entries" takedown before the reader ever reaches the caveat.
Fix
Insert one sentence immediately after The Solution (after line 20):
Done when