Ah, I think I was not terribly clear when making my point. I don't think that candidate scores are static and, even if they are, there's a significant amount of measurement error no matter the interview process.

Ideally I want a hiring "fun"-nel where a 6.2 can learn the skills necessary to become an 8.3 (or show that they've been an 8.3 all along) through exercises or reading assignments. Sort of an external training/hiring funnel. If nothing else it may also help as a way to build internal training processes and advertise the to the world some of the neat stuff you are working on. I think this is how Matasano recruits and it makes a lot of sense to me.

Pretty much no matter what there is a learning curve associate d with bringing on a new hire. Finding candidates that will attack those curves with gusto is key to building good teams, I think at least. Technical skills can be taught and refined, gumption is a bit harder to instill.

It's difficult to know when to invest the time in giving candidates feedback. While I would love to do it every time, giving any feedback during the hiring process is a legal minefield and not every candidate appreciates it. In my humble opinion, any engineer who has top-tier potential will be able to derive from the questions asked during the interview process what they need to learn in able to get a job (note, I did say "get a job", not "be a great engineer" - thinking that those are the same skillset seems to be the first mistake many job-hunters make).

It's no surprise that fresh graduates don't have an extremely clear idea about what employers in their field desire until multiple interviews later.

If potential employers divulge shortcomings in the interview then they may be slammed with a discrimination lawsuit.

If potential employers keep their mouths shut, then the overall incoming workforce takes longer to understand the rules of the hiring game.

The error bars are massive.

At the CV stage you might be a 6.2, but the error bars on whether you're a good hire might be +/- 5.8.

As you progress you're not just filtering, you're also trying to reduce the massive degree of error.

I am speaking in terms of accurate measurements...an 8.3 is actually 8.3. Precision is a different issue, but if there are no actual 8.3's in the funnel, then while filtering may improve precision there will never be an actual 8.3 at the narrow end.