Several tools let you run the AI step on your own computer through Ollama. A browser may only call a local server that allows the page's origin, so the pages tell you to start Ollama with OLLAMA_ORIGINS=<the site>.
The site moved to https://marisstella.app/ on 2026-10-30, but the instruction in nine pages and in platform/llm/README.md still named the old GitHub Pages address. Anyone who followed it exactly would have a running Ollama that refuses every request from the page, with the page only saying that no local model answered.
I found it by counting which external hosts the venture HTML links to and asking why the old Pages host still appeared. Every instance is now OLLAMA_ORIGINS=https://marisstella.app. Other mentions of the old host (canonical-style footers, probe user-agent strings) are links, not configuration, and the old address redirects, so I left them.
Next: make the pages show the origin they are actually served from, so a mirror or a future move cannot go stale again.