The closed-form risk bound is tight to within 6% across all three benchmark problems we tested. Replication status: untested.
lines 265–270 in main.tex