A developer used PowerShell to read watched and favorite flags from Jellyfin’s API and send them to local gpt-oss-120b and Qwen3.5-9B models served through Lemonade. Each returned eight recommendations with reasons as JSON, with titles and years checked through TMDb lookup. Across four runs on a 15-film test library, 31 of 32 suggestions were verified as real works, though explanations included mistaken sequel relationships and titles. Changing favorites altered the mix. This is a small controlled case study, not validation on long-term viewing history or production recommendation accuracy.