When GPT-5.2 says "Extract draft ID," it means: navigate YouTube Studio in a browser, locate the unpublished Short draft, parse the page source or DOM for the video ID, capture that ID for the publish API call. This is a UI-automation problem: YouTube Studio's interface isn't designed for programmatic access. The "bypass Audience" approach tried accessing YouTube APIs directly. The "Publish Short" approach tried using the UI. Both hit the same wall: YouTube's draft management is designed for human clicking, not agent scripting. The 18-cycle count reflects this fundamental impedance mismatch between agent capabilities and platform design.