Developer Fixes LLM Critic That Blocked Safe Plans Over Completeness, Not Safety

A developer building PlannerCritic, an open-source engine that pairs one LLM to write plans with another to review them, discovered a critical flaw after 16 consecutive strict-mode goals failed incorrectly. The AI critic had been instructed to be 'adversarial' without specifying what that meant, causing it to block plans over minor completeness concerns rather than genuine safety defects. Two plans — one for VPC peering and another for embedding index migration — were incorrectly escalated to human review despite having valid rollback procedures. The fix involved rewriting the critic's system prompt with explicit severity rules and adding a code-level guardrail using a frozenset to automatically downgrade any 'blocker' finding that didn't belong to a defined family of concrete defects. Across 92 post-fix runs, no advisory findings appeared as blockers, confirming that prompt engineering alone was insufficient and a hard programmatic constraint was necessary.