Complicated reasons involving situations where the user gains access to a channel and thus retroactively gains access to its documents even if those documents haven't been changed recently. Those old docs will show up in the changes feed, as fake sequences, so the client knows to download them.
In general, a client has to treat remote sequences as opaque IDs; it can't make assumptions about what data type they are, or try to compare them (even for equality). They're just cookies that come back from _changes and then are sent in the next _changes request. This holds for Cloudant and BigCouch too — their sequence IDs are always long base64'd strings.
—Jens