HTMLCure: Browser-Based Framework for Interactive HTML Repair

HTMLCure functions as a framework for browser experiences, assessing HTML pages following user interactions to tackle failures that arise during scrolling, hovering, clicking, resizing, or gameplay. Unlike traditional methods that rely on screenshots, this framework evaluates pages across various viewports and interaction states, capturing deterministic evidence from the browser. It also offers a VLM with curated keyframes from the executed paths. A closed-loop repair engine identifies problems, chooses state-specific repair families, tests candidates, and produces quality-approved pages for SFT. From a corpus of 97K prompts, HTMLCure increased the usable seed to a candidate pool of 63,703 quality-cleared pages.