Can I answer my own question?
I would start with 25,000-50,000 candidates that were a statistically accurate sample of the overall young population in terms of socio-economic status, geographical distribution, etc. This is not for the sake of “fairness” but rather because I am humble and do not think I could pick desired characteristics up front with high confidence.
I would assess candidates with computer exercises and then high-end motion simulators. The activity here would NOT consist of playing driving or racing games. I would try to test learning, adaptability, coordination, response time, multitasking, intelligence, etc.
When I had culled the field down in stages to around 5 candidates then I would begin actual driving. Ugly cars that go really fast can be made cheaply so I would avoid sexy/high-profit-margin series like GP2/3. I would set up a mini-academy for a year with 3 or 5 ugly/wicked-quick cars.
Candidates would be evaluated with various good and bad setups. Students would be encouraged to develop their ability to give feedback and improve the cars with engineers, but I would not directly assess this— it would simply improve their car performance over the course of the year.
I hate to say it, but the final evaluation emphasis would be on objective raw quickness and car control rather than “racecraft” and passing ability. If you are really 99.999th percentile for the former then you will be at least 95th percentile for the latter, and Sebastian Vettel has demonstrated that is good enough.
I think spending $10 million on this program would have a higher probability of discovering the next Hamilton than spending $10 million on one promising driver who has already paid to enter the traditional talent pool. My initial talent pool is 100 times as large, and surely my selection process is at least 1% as good!!