Having web documents not render in case of errors is already bad. But we already have "auto-correction" for that case - it's how HTML rendering already works in browsers.
Having an LLM hallucinate a new page in case of errors isn't a better solution, it's qualitatively worse. If you want web documents to render with errors, just use HTML.
In this brave new world we can try again. This time, though, when a parser error occurs we can spin up an Agent in the background to fix the document, looping until it passes the parser's validation, then display that! We can then have the browser automatically submit a PR or bug report to the website operator with the fix.
That way we can achieve well-defined wire formats with deterministic rendering behavior!