teacher-context token nuclei: unguided_train500_incorrect_teacher ({'n_rollouts': 2949, 'n_tokens': 1170873, 'singleton_count': 941639, 'singleton_frac': 0.8042195865819777})

token nuclei: unguided_train500_incorrect ({'n_rollouts': 2949, 'n_tokens': 1170873, 'singleton_count': 954904, 'singleton_frac': 0.8155487401280924})

token_nuclei: neg_reinforce_screen_token_nuclei -> single nuclei.parquet (684 shards merged)

token_nuclei: disjoint_rescue_token_nuclei -> single nuclei.parquet (31 shards merged)

token_nuclei: oat-zero__forced_openers_d5_k3_token_nuclei -> single nuclei.parquet (50 shards merged)

token_nuclei: base__forced_openers_d3_k3_token_nuclei -> single nuclei.parquet (48 shards merged)

oat-zero: opening-trees + forced/guided opener generations (preserve from GPU box) (part 3)

oat-zero: opening-trees + forced/guided opener generations (preserve from GPU box) (part 2)

oat-zero: opening-trees + forced/guided opener generations (preserve from GPU box)

grade_shards: 10782 rollouts / 192 problems

grade_shards: 10863 rollouts / 199 problems

grade_shards: 2538 rollouts / 199 problems

grade_shards: 10782 rollouts / 192 problems

cross-model token nuclei: Oat rollouts / base nuclei

cross-model token nuclei: Oat rollouts / base nuclei

finished (eos) leaves: qwen2.5-math-7b-oat-zero ['opening_trees_d5']

grade_shards: 10881 rollouts / 199 problems

opening_trees: stage-1 build

oat forced_openers_d3_k3 pool (504 retry)

grade_shards: 10782 rollouts / 192 problems

oat-zero opening trees: qwen2.5-math-7b-oat-zero/opening_trees_d3_val

oat-zero opening trees: qwen2.5-math-7b-oat-zero/opening_trees_d3_train

oat-zero opening trees: qwen2.5-math-7b/opening_trees_d3_val

oat-zero opening trees: qwen2.5-math-7b/opening_trees_d3_train

opening_trees: stage-1 build

opening_trees: stage-1 build

opening_trees: stage-1 build

opening_trees: stage-1 build

opening sweep prompts: -> 11 problems

opening sweep: -> 11 problems

opening-branch sweep: student prompts (6 problems)

opening-branch sweep: per-opener fix-rate (6 problems, K=16)

neg-reinforce before/after dist tables (headline lr/steps)

disjoint-rescue: top-10 branch-token sweep rollouts (POOL_SCHEMA, 640 rows, K=16)

disjoint-rescue: top-10 branch-token fix-rate (4 cases, K=16)

disjoint-rescue: fix-rate control rollouts (POOL_SCHEMA, 992 rows, both arms, K=16)

disjoint-rescue: student-vs-teacher fix-rate control (K=16, both arms)

disjoint-rescue: fix-rate control rollouts (POOL_SCHEMA, 992 rows, both arms, K=16)

disjoint-rescue: student-vs-teacher fix-rate control (K=16, both arms)

disjoint-rescue: student-vs-teacher fix-rate control (K=4, both arms)

disjoint-rescue viz nucleus store (+logZ, ours): disjoint_rescue

preserve disjoint_temp_rescue finding data

add generations/qwen3-8b-base/neg_reinforce_screen_token_nuclei

add generations/qwen3-8b-base/neg_reinforce_screen