group sample + reference
prompt learner: "i want one coffee please"
response 0 model sample
"nice try! try: 'i'd like a coffee, please.'"
response 1 model sample
"✨ great start! 👏 try — 'i'd like a coffee' — 🎯"
response 2 · reference ★ blind to judge
"close! we usually say 'i'd like a coffee, please.' want to try again?"
↑ the original tutor turn from the trace
slotted in at a random position · judge sees it as just another response · its rank becomes the reward anchor
response 3 model sample
"ok."
judge ranking → rewards
judge rank all 4 on one rubric · ties allowed
tier 1 · above ref [ 0 ] 1.0
tier 2 · the reference [ ★ ref ] 0.5
tier 3 · below ref [ 1 ] 0.15
tier 4 · below ref [ 3 ] 0.0
above ref → (0.5, 1.0]  ·  tied → 0.5  ·  below ref → [0, 0.3) (discontinuous drop from 0.5)