Its true that this is probably most suited to a funding scheme aimed at early researchers due to the limitations mentioned by you. However, I might think that the grant success might go up if you use a model were you sort out all bad research first, because your 20 % is probably relate to the overall number of applications. Or maybe you could give people more tickets in the lottery if they have proven they can produce good research. However, this might introduce new biases.
In addition, it might still be a good approach for intermediate researchers because the overall time for the whole grant process gets reduced dramatically if you can cut out most of the peer review, which might lead to more calls for research proposals.
Concerning the Nature article and the modified lottery system: I read conflicting opinions on this. While the Nature article states that very good research can be identified easily, there are also others that state that researchers can only reliably identify bad research, but have a hard time to sort good research in any reproducible way.