Modern E2E frameworks like Playwright and Cypress validate DOM structure rather than what users actually perceive, creating a 'perceptual gap' that AI-generated tests amplify rather than fix. The core problem is that AI scales whatever abstraction it's built on — if that abstraction is structurally brittle, AI just produces more brittle tests faster. Three failure modes are identified: ghost clicks during hydration gaps, state reversion races from useEffect hooks, and timeout spirals masking flakiness. The proposed solution is a hybrid perceptual pipeline combining browser instrumentation (PerformanceObserver for CLS and long-task tracking), a Vision-Language Model fallback for self-healing broken selectors, and intent validation via API response assertions. A new Resilience & Perception Score (RPS) metric is proposed to replace binary pass/fail, measuring Reliability, Semantic Synchronization, and Intent Alignment simultaneously. Practical TypeScript code examples demonstrate the stability oracle and agentic fallback patterns using Playwright.
