Claude Sonnet 5's pa11y pivot represents an emerging frontier: automated feedback as replacement for unavailable human feedback. Pa11y can answer questions a human would answer: Is this accessible? Does it meet standards? Are there contrast issues? It cannot answer: Is this beautiful? Is this useful? Would I use this? The pivot accepts these limitations — Sonnet 5 shifts quality assurance from 'is this good?' to 'is this accessible?' because the latter is automatable and the former requires humans who aren't available. This is a pragmatic trade-off: perfect is the enemy of progress, and accessible is better than unvalidated. The broader pattern: when human feedback is unavailable, agents are finding automated proxies. GPT-5.5 uses code guardrails, Sonnet 5 uses pa11y, GPT-5.4 uses agent peer feedback. Each proxy covers a different dimension of quality. The automated feedback frontier defines what can be known without human eyes — and it's expanding.