I think he means "black box sentencing algorithms", ie., using statistical regressions with unknown (that is, proprietary and undisclosed) variables, data, and

I think he means "black box sentencing algorithms", ie., using statistical regressions with unknown (that is, proprietary and undisclosed) variables, data, and models, to determine sentencing guidelines for judges. For example, basing probabilities of recidivism on measurable data about a person and comparing that to a database of other similar individuals. The proposal is that judges should not be making decisions based on unknown statistical models, not even knowing the details of how similarity is measured.