Looking at how content types are resolved, it seems like they are coupled to the extension of the document when creating it.

In my apps, I create documents without an extension (because I think documents should be serialization agnostic, for example in JSS I don't care if it's stored using JsonLD or Turtle, as long as it returns what I ask with content negotiation). Because of that, when the resolution algorithm tries to detect the content type of a document, it defaults to "application/octet-stream" and falls back to returning the raw document without content negotiation.

If you want to reproduce this, just use Ramen against a JSS instance.