Not to discredit the very capable developers discussing this, but in the interest for giving credit where credit is due, didn't Peter Todd suggest this back in his Bitcoin 2013 presentation on off-chain transactions? I seem to remember him explaining something similar on a rooftop patio in Toronto last spring after a Bitcoin Toronto meetup.
I believe this was most extensively discussed as part of a long chat that Peter Todd was a part of, so no surprise that you've seen him talk about it. Off-chain banks stuff has been a long term pet interest of his.
In that discussion we applied a merkel-sum tree data-structure— a pet datastructure that I'd previously proposed for making compact proofs of blockchain invalidity in Bitcoin (in order to make a future bitcoin world where no one runs full nodes safe from inflation and theft by miners)— to PT's bank fraud proofing application.
Search for "auditable off-chain transactions" and "Merkle-sum-tree"
(I left in a lot of unrelated stuff since it makes the meandering conversation make a bit more sense. Though a lot of this continues a long running dialog about cryptographic-wankery that has been going on for years)
Ultimately these schemes require the use of a jamming free broadcast network of some kind... otherwise they run into the same problems certificate transparency has where you can substitute the commitment on the fly. Fortunately, Bitcoin provides a global consensus mechanism which could be used to directly attach the commitment to the coins being spoken for.
I'm pretty sure I remember either Gregory Maxwell or Andrew Miller suggesting it to me first, and I think it might be the latter's idea originally. (at least in the Bitcoin space) Andrew has done a lot of work on "merkelizing" data structures: http://www.cs.umd.edu/~amiller/gpads/
EDIT: http://www.youtube.com/watch?v=4d3LA8KpdMQ#t=6m45s