This is the same problem google faced and 'solved' by with the page-rank algorithm. I've seen some attempts at applying this to the reference graph, but it doesn't seem to be taking off.

It's not taking off because, out of the box, it can't replace the gatekeepers: there needs to be a certain time period, sometimes decades, from publication until the moment a major work is recognized and cited. Cites form a DAG (old papers never cite newer ones) so do not converge on a recursive application of PageRank, newest works have all identical weights by definition.

So in the short run, you still need a prospective measure of scientific value, and acceptance for publication in a selective journal ("impact factor") serves that purpose well.

Ah of-course. I forgot the convergence requirements for pagerank. Indeed, the fact that cites are a DAG really does kind of screw with pagerank.

Note that my supposed idea was to use 'pagerank' to replace reference count as the value of a work. Not to replace peer review.