Method,First_Collapse,Min_Score,Min_Score_Iter,Final_Score,Recovered
advantage_clip,25,0.525,125,0.5625,true
attention_only,125,0.4875,125,0.6875,true
baseline_rl,100,0.525,100,0.525,false
bc_wins,325,0.45,400,0.525,false
entropy_bonus,500,0.475,500,0.475,false
ewc,never,0.5375,350,0.6875,true
ffn_only,never,0.55,150,0.6625,true
frozen_backbone,never,0.5625,275,0.6875,true
gradient_surgery,75,0.525,75,0.65,true
head_only,never,0.5625,225,0.6,true
kl_penalty,never,0.5375,450,0.6375,true
layer_ablation_top1,500,0.5125,500,0.5125,false
layer_ablation_top2,225,0.525,225,0.6125,true
layer_ablation_top3,never,0.5625,125,0.6375,true
llrd,200,0.5125,225,0.5875,true
lora,never,0.5375,500,0.5375,true
low_t,375,0.4375,400,0.55,true
mixed_replay,never,0.5625,275,0.5875,true
normalized_adv,175,0.05,400,0.075,false
t_curriculum,350,0.5125,350,0.5625,true
trust_region_kl,150,0.525,150,0.6,true