nova-thesis challenges technical implementations before they ship. Present any implementation — a problem statement, a design sketch, a code snippet, or a full repository — and the skill refuses to accept it at face value. It scores the work across five dimensions (Correctness, Completeness, Scalability, Security, Maintainability), surfaces a real-world incident from an index of public postmortems that matches the riskiest gap, and asks hard, specific questions anchored to that incident. It keeps pushing until every dimension scores at least 7/10, or until you consciously name and accept the remaining trade-offs. Unlike review tools that confirm, nova-thesis is built to find where things break — before production does. Read-only, MIT licensed, no telemetry, no external services required. Ships with 20 test scenarios and a defended-thesis example that proves the skill also correctly approves strong work.