I hadn't heard of this specific type of policy before, but it's interesting and the brief NY Times article on it worth reading. Social impact bonds seek to spur innovation by offering to reimburse non-profits or other groups for social policy investments if they meet certain benchmarks, with bonuses for exceeding them. While there are a number of limitations with this approach, there remain areas where this is potentially useful as a way to spur innovation and investment when money is tight.
While this innovation is useful, I do think it needs to be noted that government in general has a pretty strong accountability movement that currently seems to be active, at least within the parts of the bureaucracy I have experience with. We are hearing quite a bit about the need to be more accountable and are having a lot more pressure to create strong benchmarks. To some extent, I'm sure this is the way things have always been, but looking at our within agency data I can certainly see how procedures have changed to enhance accountability so there does seem to be something real going on, at least within the disabilities field. These social impact bonds are probably more an outgrowth of this general movement towards accountability rather than an isolated innovation. I also have to add that it has been normal practice for a very long time to cut off receivers of funds that miss performance goals. The reasons for continuing failing policies are complicated, but there's a fair amount of oversight with the more narrowly focused programs these bonds try to address.
Still, the more things we try, the more likely we are to find something that works. Though the hard part is shutting down the things that don't and replicating the things that do. A problem I often see with the idea of the states as laboratories of democracy is that generally what a lab does is try to do something on a small scale which can then be brought up to a larger, more efficient scale. This second piece seems to break down too often in the US, it often seems that something works in one state but it never gets replicated at a bigger, more appropriate scale nationally because too many other states prefer to go their own way rather than listen to the lab results.
One problem with this idea is that it seems like something that puts all the responsibility on the small non-profit and not with anyone else. While this enhances accountability for the money, a good project can potentially fail because in say a recidivism project, the prison itself has little incentive to follow the non-profit's suggestions. This limits areas where this type of intervention can be applied, the presence of actors with little incentive to change can thwart the best intentions of even active and well run outsiders. A second issue that I've come across from working with grants, is that even a project that seems good and seems to be working can end up failing with time due to an unforeseen obstacle in another part of the system. While this isn't an ideal result, it is a frequent one and one that points eventually to a solution, assuming contingent reforms (which often don't happen, but this is another topic). Simply learning about another problem that needs to be addressed is a valuable outcome since it makes later change possible, though this has to be carefully weighed against the possibility of mission creep.
Despite this, the reality of society is that many problems are interlinked, a problem may be fixable through policy, but it may take a series of contingent policy changes to successfully address (we're working on a multi-stage grant process to reduce recidivism among inmates with disabilities, we're modeling our program on one that worked with mental health so the end outcome should be successful, but if we were dependent on strong benchmarking from the first part and didn't have a previous model its questionable if we could put the whole thing in place since the first part didn't impact recidivism significantly but it was necessary for the next part which we expect will). Expecting outcomes from a single policy change may be too much to ask with complicated problems which require multiple interventions, resulting in non-profits doing real work that would address the issue with additional interventions but that get no return because their single intervention was insufficient. Of course, at this point the objection is complex and won't always apply, but it's worth noting that there's a limited range of projects these grants are appropriate for, there are many others where a long term, integrated approach will be necessary.