Web4AI: Intent-Driven Evolutionary Self-Adaptation of Web Applications with Large Language Models

LLMs combined with evolutionary search can discover improved programs when a precise fitness function is available, but web applications rarely have one: their goals are stated in business terms, trade off against each other, and are bounded by requirements that must never be violated. We present Web4AI, in which operators declare objectives and constraints in a small intent language; a compiler translates each intent into a constrained multi-objective fitness specification; an LLM-guided evolutionary engine proposes code variants; and a safety envelope of tests, static checks, staged rollout with non-inferiority testing, and automatic rollback governs what reaches users. We define the language's syntax and semantics, map it onto the MAPE-K self-adaptation loop, analyze risks such as metric gaming and evolutionary drift, and specify an evaluation protocol.