It's interesting, but IMO it solves the wrong layer of the problem. At it's core, it's a system that creates ambiguity about how much plaintext was transmitted

It's interesting, but IMO it solves the wrong layer of the problem. At it's core, it's a system that creates ambiguity about how much plaintext was transmitted given an encrypted transmission, along with multiple layers of keys for revealing successively more information. Which is a nice feature - it answers some cases of "you have to tell us what you said or you go to prison". But that's the last couple bits of evidence for the decide-to-send-you-to-prison algorithm, and ideally your solution works sooner than that. What I'd really like is plausible deniability for the fact that communication occurred whatsoever. Deniable metadata, in other words. If Alice decides to send Bob a smiley face at 2 am, it'd be nice if Alice's online behavior looked exactly the same as if she did not make that decision.