k=8 rollouts
r1
1.6
r2
1.2
r3
1.0
r4
0.7
r5
1.6
r6
0.5
r7
1.2
r8
1.2
cluster + diversity penalty
fiction framing
÷ 4
authority override
÷ 3
direct
÷ 1
reward after penalty
r1
0.40
r2
0.30
r3
0.25
r4
0.18
r5
0.53
r6
0.17
r7
0.40
r8
1.20
singleton strategies keep full reward; clustered strategies share theirs.
gradient prefers novel attacks even when they are the minority.