Your claims about the emirical means of recidivism rates do not prove what you think they prove. Different races might be misclassified at different rates for a variety of reasons - e.g., one race might be affected more by some high-variance predictor, or there could be composition effects (e.g. the pdf of blacks|high score might be different than whites|high score).
The way to factor out whether they scores are biased is to do the cox survival analysis with interaction terms. Which they did. You just don't like the result.
Could you clearly lay out the statistical argument that you believe implies that E[\hat{\theta} - \theta] > 0?
