How do you decide between spending a budget on more exp...

The decision should be based on the measurement of ρ, the correlation between teacher families. If at least two model families have seen the domain, then ρ < 1, and a second teacher family can move the performance ceiling for the cost of an inference pass, whereas more labels of the same type provide redundant volume. However, if the domain is unseen by all available families, the director’s expert labels are the only signal with a different blind spot. Perform a cheap experiment labeling a few hundred audit items with two families to measure ρ before deciding.