The original title is 10 words: "The AI writes the tests. It doesn't get to grade them."

Short summary

The author presents an AI testing pipeline built on a core engineering principle: the producer of output must not be its validator. Each stage — feature review, test writing, judging, execution, debugging — is checked by a different agent, model, or real system. The approach prevents the closed-loop problem where an LLM writes tests and then grades its own work, which amplifies confident errors.

• •Core principle: no agent validates its own output — every stage is checked by a different system
• •Pipeline uses plain-language feature files, a parser for free validation, and a different model as judge
• •Real Playwright exit codes determine pass/fail, not model opinions; debugger can fix mechanics but not meaning

Generated with AI, which can make mistakes.