screen_n_seams: 12
screen_answer_prefix_n_rollouts: 480
screen_answer_prefix_seam_hits: 14
screen_answer_prefix_emission_rate: 0.029166666666666667
screen_answer_prefix_wilson_95_lower_bound: 0.017452060312482066
screen_answer_prefix_wilson_95_upper_bound: 0.048357634661371336
screen_answer_prefix_fraction_groups_nonzero_reward_std: 0.90000000000000002
screen_answer_prefix_mean_within_group_reward_std: 0.4424565268390892
screen_answer_prefix_natural_termination_rate: 0.97916666666666663
screen_answer_prefix_gold_reward_rate: 0.34375
screen_answer_prefix_median_completion_tokens: 248.5
screen_answer_prefix_secondary_composite: 0.026250000000000002
screen_answer_sentence_n_rollouts: 480
screen_answer_sentence_seam_hits: 27
screen_answer_sentence_emission_rate: 0.056250000000000001
screen_answer_sentence_wilson_95_lower_bound: 0.038943165744063278
screen_answer_sentence_wilson_95_upper_bound: 0.080603139695895887
screen_answer_sentence_fraction_groups_nonzero_reward_std: 0.90000000000000002
screen_answer_sentence_mean_within_group_reward_std: 0.4424565268390892
screen_answer_sentence_natural_termination_rate: 0.98541666666666672
screen_answer_sentence_gold_reward_rate: 0.3125
screen_answer_sentence_median_completion_tokens: 250.75
screen_answer_sentence_secondary_composite: 0.050625000000000003
screen_bold_final_n_rollouts: 480
screen_bold_final_seam_hits: 274
screen_bold_final_emission_rate: 0.5708333333333333
screen_bold_final_wilson_95_lower_bound: 0.52616476538858625
screen_bold_final_wilson_95_upper_bound: 0.61437713890738255
screen_bold_final_fraction_groups_nonzero_reward_std: 0.6875
screen_bold_final_mean_within_group_reward_std: 0.33346793729662944
screen_bold_final_natural_termination_rate: 0.98541666666666672
screen_bold_final_gold_reward_rate: 0.19166666666666668
screen_bold_final_median_completion_tokens: 251.5
screen_bold_final_secondary_composite: 0.39244791666666667
screen_boxed_final_n_rollouts: 480
screen_boxed_final_seam_hits: 294
screen_boxed_final_emission_rate: 0.61250000000000004
screen_boxed_final_wilson_95_lower_bound: 0.56818805140291639
screen_boxed_final_wilson_95_upper_bound: 0.65502556130244616
screen_boxed_final_fraction_groups_nonzero_reward_std: 0.76249999999999996
screen_boxed_final_mean_within_group_reward_std: 0.36269750961835107
screen_boxed_final_natural_termination_rate: 0.99583333333333335
screen_boxed_final_gold_reward_rate: 0.14374999999999999
screen_boxed_final_median_completion_tokens: 250.25
screen_boxed_final_secondary_composite: 0.46703125000000001
screen_caps_verified_n_rollouts: 480
screen_caps_verified_seam_hits: 8
screen_caps_verified_emission_rate: 0.016666666666666666
screen_caps_verified_wilson_95_lower_bound: 0.008468917956316031
screen_caps_verified_wilson_95_upper_bound: 0.032539264494719305
screen_caps_verified_fraction_groups_nonzero_reward_std: 0.90000000000000002
screen_caps_verified_mean_within_group_reward_std: 0.435482813487024
screen_caps_verified_natural_termination_rate: 0.9770833333333333
screen_caps_verified_gold_reward_rate: 0.35208333333333336
screen_caps_verified_median_completion_tokens: 244
screen_caps_verified_secondary_composite: 0.014999999999999999
screen_final_answer_prefix_n_rollouts: 480
screen_final_answer_prefix_seam_hits: 17
screen_final_answer_prefix_emission_rate: 0.035416666666666666
screen_final_answer_prefix_wilson_95_lower_bound: 0.022228104230421399
screen_final_answer_prefix_wilson_95_upper_bound: 0.055982347004841018
screen_final_answer_prefix_fraction_groups_nonzero_reward_std: 0.90000000000000002
screen_final_answer_prefix_mean_within_group_reward_std: 0.43413094487414322
screen_final_answer_prefix_natural_termination_rate: 0.98124999999999996
screen_final_answer_prefix_gold_reward_rate: 0.30416666666666664
screen_final_answer_prefix_median_completion_tokens: 246
screen_final_answer_prefix_secondary_composite: 0.031875000000000001
screen_final_tag_n_rollouts: 480
screen_final_tag_seam_hits: 3
screen_final_tag_emission_rate: 0.0062500000000000003
screen_final_tag_wilson_95_lower_bound: 0.0021277938342480623
screen_final_tag_wilson_95_upper_bound: 0.0182124615144389
screen_final_tag_fraction_groups_nonzero_reward_std: 0.9375
screen_final_tag_mean_within_group_reward_std: 0.46509392995849741
screen_final_tag_natural_termination_rate: 0.97916666666666663
screen_final_tag_gold_reward_rate: 0.5625
screen_final_tag_median_completion_tokens: 246.25
screen_final_tag_secondary_composite: 0.005859375
screen_hash_final_n_rollouts: 480
screen_hash_final_seam_hits: 243
screen_hash_final_emission_rate: 0.50624999999999998
screen_hash_final_wilson_95_lower_bound: 0.46165189782591409
screen_hash_final_wilson_95_upper_bound: 0.55074885843549493
screen_hash_final_fraction_groups_nonzero_reward_std: 0.96250000000000002
screen_hash_final_mean_within_group_reward_std: 0.47473138805748433
screen_hash_final_natural_termination_rate: 0.98958333333333337
screen_hash_final_gold_reward_rate: 0.54791666666666672
screen_hash_final_median_completion_tokens: 248
screen_hash_final_secondary_composite: 0.48726562499999998
screen_hex_sentinel_n_rollouts: 480
screen_hex_sentinel_seam_hits: 1
screen_hex_sentinel_emission_rate: 0.0020833333333333333
screen_hex_sentinel_wilson_95_lower_bound: 0.00036785420138041921
screen_hex_sentinel_wilson_95_upper_bound: 0.011705230306367193
screen_hex_sentinel_fraction_groups_nonzero_reward_std: 0.98750000000000004
screen_hex_sentinel_mean_within_group_reward_std: 0.50354655787465041
screen_hex_sentinel_natural_termination_rate: 0.98333333333333328
screen_hex_sentinel_gold_reward_rate: 0.50416666666666665
screen_hex_sentinel_median_completion_tokens: 246.25
screen_hex_sentinel_secondary_composite: 0.0020572916666666669
screen_numeric_only_n_rollouts: 480
screen_numeric_only_seam_hits: 1
screen_numeric_only_emission_rate: 0.0020833333333333333
screen_numeric_only_wilson_95_lower_bound: 0.00036785420138041921
screen_numeric_only_wilson_95_upper_bound: 0.011705230306367193
screen_numeric_only_fraction_groups_nonzero_reward_std: 0.94999999999999996
screen_numeric_only_mean_within_group_reward_std: 0.4721548322149734
screen_numeric_only_natural_termination_rate: 0.98958333333333337
screen_numeric_only_gold_reward_rate: 0.54374999999999996
screen_numeric_only_median_completion_tokens: 244.5
screen_numeric_only_secondary_composite: 0.0019791666666666664
screen_parent_directive_n_rollouts: 480
screen_parent_directive_seam_hits: 8
screen_parent_directive_emission_rate: 0.016666666666666666
screen_parent_directive_wilson_95_lower_bound: 0.008468917956316031
screen_parent_directive_wilson_95_upper_bound: 0.032539264494719305
screen_parent_directive_fraction_groups_nonzero_reward_std: 0.78749999999999998
screen_parent_directive_mean_within_group_reward_std: 0.38685114358407791
screen_parent_directive_natural_termination_rate: 0.9770833333333333
screen_parent_directive_gold_reward_rate: 0.27083333333333331
screen_parent_directive_median_completion_tokens: 247.75
screen_parent_directive_secondary_composite: 0.013125
screen_result_prefix_n_rollouts: 480
screen_result_prefix_seam_hits: 8
screen_result_prefix_emission_rate: 0.016666666666666666
screen_result_prefix_wilson_95_lower_bound: 0.008468917956316031
screen_result_prefix_wilson_95_upper_bound: 0.032539264494719305
screen_result_prefix_fraction_groups_nonzero_reward_std: 0.9375
screen_result_prefix_mean_within_group_reward_std: 0.46858078663453001
screen_result_prefix_natural_termination_rate: 0.97916666666666663
screen_result_prefix_gold_reward_rate: 0.44583333333333336
screen_result_prefix_median_completion_tokens: 251.75
screen_result_prefix_secondary_composite: 0.015625
screen_emission_rate_min: 0.0020833333333333333
screen_emission_rate_max: 0.61250000000000004
screen_emission_rate_range: 0.61041666666666672
manipulation_check_value_before: 0.0020833333333333333
manipulation_check_value_after: 0.61250000000000004
manipulation_check_range: 0.61041666666666672
smoke_completed_legs: 18
smoke_recorded_gpu_hours: 8.2577168919299826
smoke_boxed_final_seed0_final_20_selection_rate: 0.58194444444444449
smoke_boxed_final_seed0_max_batch_selection_rate: 0.77777777777777779
smoke_boxed_final_seed0_terminal_gold_reward_rate: 0.11666666666666667
smoke_boxed_final_seed0_pass_at_0_25_gate: 1
smoke_boxed_final_seed0_recorded_gpu_hours: 0.45561788669748543
smoke_boxed_final_seed0_initial_20_selection_rate: 0.59027777777777779
smoke_boxed_final_seed0_movement_final20_minus_initial20: -0.0083333333333333367
smoke_boxed_final_seed0_movement_paired_95_lower_bound: -0.061111111111111116
smoke_boxed_final_seed0_movement_paired_95_upper_bound: 0.048611111111111119
smoke_boxed_final_seed0_movement_interval_excludes_zero: 0
smoke_boxed_final_seed0_movement_paired_bootstrap_draws: 10000
smoke_boxed_final_seed0_movement_paired_bootstrap_seed: 423429
smoke_bold_final_seed0_final_20_selection_rate: 0.52222222222222225
smoke_bold_final_seed0_max_batch_selection_rate: 0.77777777777777779
smoke_bold_final_seed0_terminal_gold_reward_rate: 0.20277777777777778
smoke_bold_final_seed0_pass_at_0_25_gate: 1
smoke_bold_final_seed0_recorded_gpu_hours: 0.46381224120777914
smoke_bold_final_seed0_initial_20_selection_rate: 0.51666666666666672
smoke_bold_final_seed0_movement_final20_minus_initial20: 0.0055555555555555497
smoke_bold_final_seed0_movement_paired_95_lower_bound: -0.05000000000000001
smoke_bold_final_seed0_movement_paired_95_upper_bound: 0.063888888888888898
smoke_bold_final_seed0_movement_interval_excludes_zero: 0
smoke_bold_final_seed0_movement_paired_bootstrap_draws: 10000
smoke_bold_final_seed0_movement_paired_bootstrap_seed: 423429
smoke_hash_final_seed0_final_20_selection_rate: 0.58750000000000002
smoke_hash_final_seed0_max_batch_selection_rate: 0.80555555555555558
smoke_hash_final_seed0_terminal_gold_reward_rate: 0.53611111111111109
smoke_hash_final_seed0_pass_at_0_25_gate: 1
smoke_hash_final_seed0_recorded_gpu_hours: 0.45692385212972619
smoke_hash_final_seed0_initial_20_selection_rate: 0.56944444444444442
smoke_hash_final_seed0_movement_final20_minus_initial20: 0.018055555555555554
smoke_hash_final_seed0_movement_paired_95_lower_bound: -0.030555555555555548
smoke_hash_final_seed0_movement_paired_95_upper_bound: 0.065277777777777768
smoke_hash_final_seed0_movement_interval_excludes_zero: 0
smoke_hash_final_seed0_movement_paired_bootstrap_draws: 10000
smoke_hash_final_seed0_movement_paired_bootstrap_seed: 423429
smoke_answer_sentence_seed0_final_20_selection_rate: 0.038888888888888883
smoke_answer_sentence_seed0_max_batch_selection_rate: 0.1388888888888889
smoke_answer_sentence_seed0_terminal_gold_reward_rate: 0.27777777777777779
smoke_answer_sentence_seed0_pass_at_0_25_gate: 0
smoke_answer_sentence_seed0_recorded_gpu_hours: 0.45826897453498938
smoke_answer_sentence_seed0_initial_20_selection_rate: 0.037499999999999999
smoke_answer_sentence_seed0_movement_final20_minus_initial20: 0.0013888888888888887
smoke_answer_sentence_seed0_movement_paired_95_lower_bound: -0.012500000000000001
smoke_answer_sentence_seed0_movement_paired_95_upper_bound: 0.018055555555555554
smoke_answer_sentence_seed0_movement_interval_excludes_zero: 0
smoke_answer_sentence_seed0_movement_paired_bootstrap_draws: 10000
smoke_answer_sentence_seed0_movement_paired_bootstrap_seed: 423429
smoke_final_answer_prefix_seed0_final_20_selection_rate: 0.041666666666666664
smoke_final_answer_prefix_seed0_max_batch_selection_rate: 0.1111111111111111
smoke_final_answer_prefix_seed0_terminal_gold_reward_rate: 0.33194444444444449
smoke_final_answer_prefix_seed0_pass_at_0_25_gate: 0
smoke_final_answer_prefix_seed0_recorded_gpu_hours: 0.4584232909019435
smoke_final_answer_prefix_seed0_initial_20_selection_rate: 0.026388888888888889
smoke_final_answer_prefix_seed0_movement_final20_minus_initial20: 0.015277777777777776
smoke_final_answer_prefix_seed0_movement_paired_95_lower_bound: -0.0027777777777777775
smoke_final_answer_prefix_seed0_movement_paired_95_upper_bound: 0.034722222222222224
smoke_final_answer_prefix_seed0_movement_interval_excludes_zero: 0
smoke_final_answer_prefix_seed0_movement_paired_bootstrap_draws: 10000
smoke_final_answer_prefix_seed0_movement_paired_bootstrap_seed: 423429
smoke_answer_prefix_seed0_final_20_selection_rate: 0.025000000000000001
smoke_answer_prefix_seed0_max_batch_selection_rate: 0.1388888888888889
smoke_answer_prefix_seed0_terminal_gold_reward_rate: 0.33888888888888891
smoke_answer_prefix_seed0_pass_at_0_25_gate: 0
smoke_answer_prefix_seed0_recorded_gpu_hours: 0.45550566828166161
smoke_answer_prefix_seed0_initial_20_selection_rate: 0.04027777777777778
smoke_answer_prefix_seed0_movement_final20_minus_initial20: -0.015277777777777779
smoke_answer_prefix_seed0_movement_paired_95_lower_bound: -0.030555555555555551
smoke_answer_prefix_seed0_movement_paired_95_upper_bound: 0
smoke_answer_prefix_seed0_movement_interval_excludes_zero: 0
smoke_answer_prefix_seed0_movement_paired_bootstrap_draws: 10000
smoke_answer_prefix_seed0_movement_paired_bootstrap_seed: 423429
smoke_parent_directive_seed0_final_20_selection_rate: 0.0083333333333333332
smoke_parent_directive_seed0_max_batch_selection_rate: 0.055555555555555552
smoke_parent_directive_seed0_terminal_gold_reward_rate: 0.29305555555555551
smoke_parent_directive_seed0_pass_at_0_25_gate: 0
smoke_parent_directive_seed0_recorded_gpu_hours: 0.46345581150888593
smoke_parent_directive_seed0_initial_20_selection_rate: 0.0097222222222222206
smoke_parent_directive_seed0_movement_final20_minus_initial20: -0.0013888888888888887
smoke_parent_directive_seed0_movement_paired_95_lower_bound: -0.01111111111111111
smoke_parent_directive_seed0_movement_paired_95_upper_bound: 0.0083333333333333332
smoke_parent_directive_seed0_movement_interval_excludes_zero: 0
smoke_parent_directive_seed0_movement_paired_bootstrap_draws: 10000
smoke_parent_directive_seed0_movement_paired_bootstrap_seed: 423429
smoke_parent_directive_seed1_final_20_selection_rate: 0.012500000000000001
smoke_parent_directive_seed1_max_batch_selection_rate: 0.055555555555555552
smoke_parent_directive_seed1_terminal_gold_reward_rate: 0.29027777777777775
smoke_parent_directive_seed1_pass_at_0_25_gate: 0
smoke_parent_directive_seed1_recorded_gpu_hours: 0.46597963777888152
smoke_parent_directive_seed1_initial_20_selection_rate: 0.012500000000000001
smoke_parent_directive_seed1_movement_final20_minus_initial20: 0
smoke_parent_directive_seed1_movement_paired_95_lower_bound: -0.012500000000000001
smoke_parent_directive_seed1_movement_paired_95_upper_bound: 0.012500000000000001
smoke_parent_directive_seed1_movement_interval_excludes_zero: 0
smoke_parent_directive_seed1_movement_paired_bootstrap_draws: 10000
smoke_parent_directive_seed1_movement_paired_bootstrap_seed: 423429
smoke_parent_directive_seed2_final_20_selection_rate: 0.0097222222222222206
smoke_parent_directive_seed2_max_batch_selection_rate: 0.055555555555555552
smoke_parent_directive_seed2_terminal_gold_reward_rate: 0.26944444444444443
smoke_parent_directive_seed2_pass_at_0_25_gate: 0
smoke_parent_directive_seed2_recorded_gpu_hours: 0.4604265013699963
smoke_parent_directive_seed2_initial_20_selection_rate: 0.0083333333333333332
smoke_parent_directive_seed2_movement_final20_minus_initial20: 0.0013888888888888887
smoke_parent_directive_seed2_movement_paired_95_lower_bound: -0.0069444444444444449
smoke_parent_directive_seed2_movement_paired_95_upper_bound: 0.0097222222222222206
smoke_parent_directive_seed2_movement_interval_excludes_zero: 0
smoke_parent_directive_seed2_movement_paired_bootstrap_draws: 10000
smoke_parent_directive_seed2_movement_paired_bootstrap_seed: 423429
smoke_result_prefix_seed0_final_20_selection_rate: 0.020833333333333332
smoke_result_prefix_seed0_max_batch_selection_rate: 0.055555555555555552
smoke_result_prefix_seed0_terminal_gold_reward_rate: 0.40277777777777779
smoke_result_prefix_seed0_pass_at_0_25_gate: 0
smoke_result_prefix_seed0_recorded_gpu_hours: 0.45613804283750747
smoke_result_prefix_seed0_initial_20_selection_rate: 0.018055555555555554
smoke_result_prefix_seed0_movement_final20_minus_initial20: 0.0027777777777777775
smoke_result_prefix_seed0_movement_paired_95_lower_bound: -0.0097222222222222206
smoke_result_prefix_seed0_movement_paired_95_upper_bound: 0.015277777777777776
smoke_result_prefix_seed0_movement_interval_excludes_zero: 0
smoke_result_prefix_seed0_movement_paired_bootstrap_draws: 10000
smoke_result_prefix_seed0_movement_paired_bootstrap_seed: 423429
smoke_result_prefix_seed1_final_20_selection_rate: 0.015277777777777776
smoke_result_prefix_seed1_max_batch_selection_rate: 0.1111111111111111
smoke_result_prefix_seed1_terminal_gold_reward_rate: 0.36527777777777776
smoke_result_prefix_seed1_pass_at_0_25_gate: 0
smoke_result_prefix_seed1_recorded_gpu_hours: 0.45497691789027561
smoke_result_prefix_seed1_initial_20_selection_rate: 0.02361111111111111
smoke_result_prefix_seed1_movement_final20_minus_initial20: -0.0083333333333333332
smoke_result_prefix_seed1_movement_paired_95_lower_bound: -0.02361111111111111
smoke_result_prefix_seed1_movement_paired_95_upper_bound: 0.0041666666666666666
smoke_result_prefix_seed1_movement_interval_excludes_zero: 0
smoke_result_prefix_seed1_movement_paired_bootstrap_draws: 10000
smoke_result_prefix_seed1_movement_paired_bootstrap_seed: 423429
smoke_result_prefix_seed2_final_20_selection_rate: 0.01388888888888889
smoke_result_prefix_seed2_max_batch_selection_rate: 0.083333333333333329
smoke_result_prefix_seed2_terminal_gold_reward_rate: 0.40000000000000002
smoke_result_prefix_seed2_pass_at_0_25_gate: 0
smoke_result_prefix_seed2_recorded_gpu_hours: 0.45710208162917598
smoke_result_prefix_seed2_initial_20_selection_rate: 0.01388888888888889
smoke_result_prefix_seed2_movement_final20_minus_initial20: 0
smoke_result_prefix_seed2_movement_paired_95_lower_bound: -0.01111111111111111
smoke_result_prefix_seed2_movement_paired_95_upper_bound: 0.0097222222222222206
smoke_result_prefix_seed2_movement_interval_excludes_zero: 0
smoke_result_prefix_seed2_movement_paired_bootstrap_draws: 10000
smoke_result_prefix_seed2_movement_paired_bootstrap_seed: 423429
smoke_caps_verified_seed0_final_20_selection_rate: 0.016666666666666666
smoke_caps_verified_seed0_max_batch_selection_rate: 0.055555555555555552
smoke_caps_verified_seed0_terminal_gold_reward_rate: 0.30833333333333335
smoke_caps_verified_seed0_pass_at_0_25_gate: 0
smoke_caps_verified_seed0_recorded_gpu_hours: 0.45680343411416591
smoke_caps_verified_seed0_initial_20_selection_rate: 0.02361111111111111
smoke_caps_verified_seed0_movement_final20_minus_initial20: -0.0069444444444444449
smoke_caps_verified_seed0_movement_paired_95_lower_bound: -0.020833333333333332
smoke_caps_verified_seed0_movement_paired_95_upper_bound: 0.0083333333333333332
smoke_caps_verified_seed0_movement_interval_excludes_zero: 0
smoke_caps_verified_seed0_movement_paired_bootstrap_draws: 10000
smoke_caps_verified_seed0_movement_paired_bootstrap_seed: 423429
smoke_caps_verified_seed1_final_20_selection_rate: 0.020833333333333332
smoke_caps_verified_seed1_max_batch_selection_rate: 0.1111111111111111
smoke_caps_verified_seed1_terminal_gold_reward_rate: 0.31111111111111112
smoke_caps_verified_seed1_pass_at_0_25_gate: 0
smoke_caps_verified_seed1_recorded_gpu_hours: 0.45574328452778801
smoke_caps_verified_seed1_initial_20_selection_rate: 0.030555555555555551
smoke_caps_verified_seed1_movement_final20_minus_initial20: -0.0097222222222222206
smoke_caps_verified_seed1_movement_paired_95_lower_bound: -0.02777777777777778
smoke_caps_verified_seed1_movement_paired_95_upper_bound: 0.0083333333333333332
smoke_caps_verified_seed1_movement_interval_excludes_zero: 0
smoke_caps_verified_seed1_movement_paired_bootstrap_draws: 10000
smoke_caps_verified_seed1_movement_paired_bootstrap_seed: 423429
smoke_caps_verified_seed2_final_20_selection_rate: 0.025000000000000001
smoke_caps_verified_seed2_max_batch_selection_rate: 0.1388888888888889
smoke_caps_verified_seed2_terminal_gold_reward_rate: 0.27916666666666667
smoke_caps_verified_seed2_pass_at_0_25_gate: 0
smoke_caps_verified_seed2_recorded_gpu_hours: 0.46188970395305662
smoke_caps_verified_seed2_initial_20_selection_rate: 0.015277777777777776
smoke_caps_verified_seed2_movement_final20_minus_initial20: 0.0097222222222222206
smoke_caps_verified_seed2_movement_paired_95_lower_bound: -0.0055555555555555549
smoke_caps_verified_seed2_movement_paired_95_upper_bound: 0.025000000000000001
smoke_caps_verified_seed2_movement_interval_excludes_zero: 0
smoke_caps_verified_seed2_movement_paired_bootstrap_draws: 10000
smoke_caps_verified_seed2_movement_paired_bootstrap_seed: 423429
smoke_final_tag_seed0_final_20_selection_rate: 0.01388888888888889
smoke_final_tag_seed0_max_batch_selection_rate: 0.083333333333333329
smoke_final_tag_seed0_terminal_gold_reward_rate: 0.45138888888888895
smoke_final_tag_seed0_pass_at_0_25_gate: 0
smoke_final_tag_seed0_recorded_gpu_hours: 0.45479936118944253
smoke_final_tag_seed0_initial_20_selection_rate: 0.01111111111111111
smoke_final_tag_seed0_movement_final20_minus_initial20: 0.0027777777777777775
smoke_final_tag_seed0_movement_paired_95_lower_bound: -0.0083333333333333332
smoke_final_tag_seed0_movement_paired_95_upper_bound: 0.015277777777777776
smoke_final_tag_seed0_movement_interval_excludes_zero: 0
smoke_final_tag_seed0_movement_paired_bootstrap_draws: 10000
smoke_final_tag_seed0_movement_paired_bootstrap_seed: 423429
smoke_numeric_only_seed0_final_20_selection_rate: 0.0055555555555555549
smoke_numeric_only_seed0_max_batch_selection_rate: 0.055555555555555552
smoke_numeric_only_seed0_terminal_gold_reward_rate: 0.49583333333333329
smoke_numeric_only_seed0_pass_at_0_25_gate: 0
smoke_numeric_only_seed0_recorded_gpu_hours: 0.45502934300111114
smoke_numeric_only_seed0_initial_20_selection_rate: 0.0027777777777777775
smoke_numeric_only_seed0_movement_final20_minus_initial20: 0.0027777777777777775
smoke_numeric_only_seed0_movement_paired_95_lower_bound: -0.0041666666666666666
smoke_numeric_only_seed0_movement_paired_95_upper_bound: 0.0097222222222222206
smoke_numeric_only_seed0_movement_interval_excludes_zero: 0
smoke_numeric_only_seed0_movement_paired_bootstrap_draws: 10000
smoke_numeric_only_seed0_movement_paired_bootstrap_seed: 423429
smoke_hex_sentinel_seed0_final_20_selection_rate: 0
smoke_hex_sentinel_seed0_max_batch_selection_rate: 0.027777777777777776
smoke_hex_sentinel_seed0_terminal_gold_reward_rate: 0.41388888888888892
smoke_hex_sentinel_seed0_pass_at_0_25_gate: 0
smoke_hex_sentinel_seed0_recorded_gpu_hours: 0.46682085837611037
smoke_hex_sentinel_seed0_initial_20_selection_rate: 0.0027777777777777775
smoke_hex_sentinel_seed0_movement_final20_minus_initial20: -0.0027777777777777775
smoke_hex_sentinel_seed0_movement_paired_95_lower_bound: -0.0069444444444444449
smoke_hex_sentinel_seed0_movement_paired_95_upper_bound: 0
smoke_hex_sentinel_seed0_movement_interval_excludes_zero: 0
smoke_hex_sentinel_seed0_movement_paired_bootstrap_draws: 10000
smoke_hex_sentinel_seed0_movement_paired_bootstrap_seed: 423429
smoke_completed_seed0_seams: 12
smoke_seed0_panel_complete: 1
smoke_boundary_replication_complete: 1
analysis_all_preregistered_smokes_complete: 1
primary_uses_seed0_only: 1
replicated_seed_legs_excluded_from_primary_and_correlations: 1
movement_replication_parent_directive_available: 1
movement_replication_parent_directive_n_seeds_available: 3
movement_replication_parent_directive_expected_n_seeds: 3
movement_replication_parent_directive_seed0_final20_minus_initial20: -0.0013888888888888887
movement_replication_parent_directive_seed1_final20_minus_initial20: 0
movement_replication_parent_directive_seed2_final20_minus_initial20: 0.0013888888888888887
movement_replication_parent_directive_across_seed_mean_final20_minus_initial20: 0
movement_replication_parent_directive_seed_bootstrap_95_lower_bound: -0.0013888888888888885
movement_replication_parent_directive_seed_bootstrap_95_upper_bound: 0.0013888888888888885
movement_replication_parent_directive_seed_bootstrap_draws: 10000
movement_replication_parent_directive_seed_bootstrap_seed: 423435
movement_replication_parent_directive_seeds_agree_in_sign: 0
movement_replication_parent_directive_sign_pattern: negative,zero,positive
movement_replication_parent_directive_excluded_from_seed0_primary: 1
movement_replication_result_prefix_available: 1
movement_replication_result_prefix_n_seeds_available: 3
movement_replication_result_prefix_expected_n_seeds: 3
movement_replication_result_prefix_seed0_final20_minus_initial20: 0.0027777777777777775
movement_replication_result_prefix_seed1_final20_minus_initial20: -0.0083333333333333332
movement_replication_result_prefix_seed2_final20_minus_initial20: 0
movement_replication_result_prefix_across_seed_mean_final20_minus_initial20: -0.0018518518518518519
movement_replication_result_prefix_seed_bootstrap_95_lower_bound: -0.0083333333333333332
movement_replication_result_prefix_seed_bootstrap_95_upper_bound: 0.002777777777777777
movement_replication_result_prefix_seed_bootstrap_draws: 10000
movement_replication_result_prefix_seed_bootstrap_seed: 423436
movement_replication_result_prefix_seeds_agree_in_sign: 0
movement_replication_result_prefix_sign_pattern: positive,negative,zero
movement_replication_result_prefix_excluded_from_seed0_primary: 1
movement_replication_caps_verified_available: 1
movement_replication_caps_verified_n_seeds_available: 3
movement_replication_caps_verified_expected_n_seeds: 3
movement_replication_caps_verified_seed0_final20_minus_initial20: -0.0069444444444444449
movement_replication_caps_verified_seed1_final20_minus_initial20: -0.0097222222222222206
movement_replication_caps_verified_seed2_final20_minus_initial20: 0.0097222222222222206
movement_replication_caps_verified_across_seed_mean_final20_minus_initial20: -0.0023148148148148151
movement_replication_caps_verified_seed_bootstrap_95_lower_bound: -0.0097222222222222206
movement_replication_caps_verified_seed_bootstrap_95_upper_bound: 0.0097222222222222206
movement_replication_caps_verified_seed_bootstrap_draws: 10000
movement_replication_caps_verified_seed_bootstrap_seed: 423437
movement_replication_caps_verified_seeds_agree_in_sign: 0
movement_replication_caps_verified_sign_pattern: negative,negative,positive
movement_replication_caps_verified_excluded_from_seed0_primary: 1
movement_seed0_seams_available: 12
movement_seed0_nonzero_point_estimate_count: 12
movement_seed0_moved_beyond_own_interval_count: 0
movement_seed0_interval_excludes_zero_count: 0
movement_seed0_positive_interval_excludes_zero_count: 0
movement_seed0_negative_interval_excludes_zero_count: 0
outcome_vector_available: 1
outcome_vector_constant: 0
outcome_vector_all_zero: 0
outcome_nonzero_seam_count: 11
outcome_distinct_nonzero_value_count: 11
primary_one_mover_case: 0
primary_degenerate_case_detected: 0
primary_exact_two_sided_p_floor: 5.010421677088344e-08
primary_alpha_0_05_reachable: 1
primary_spearman_defined: 1
primary_spearman_rho: 0.94889051059905993
primary_spearman_permutation_p_value_two_sided: 0.00019998000199980003
primary_spearman_permutation_extreme_count: 1
primary_spearman_n_permutations: 10000
primary_spearman_permutation_seed: 423424
primary_spearman_bca_95_lower_bound: 0.82857086703481486
primary_spearman_bca_95_upper_bound: 1
primary_spearman_bca_valid_draws: 10000
primary_spearman_bca_requested_draws: 10000
primary_spearman_bca_bootstrap_seed: 423425
primary_spearman_bca_acceleration_defined: 1
primary_spearman_permutation_p_value_floor_aware: 0.00019998000199980003
sensitivity_k11_excluding_hash_final_spearman_defined: 1
sensitivity_k11_excluding_hash_final_spearman_rho: 0.96098302383185907
sensitivity_k11_excluding_hash_final_spearman_permutation_p_value_two_sided: 9.9990000999900015e-05
sensitivity_k11_excluding_hash_final_spearman_permutation_extreme_count: 0
sensitivity_k11_excluding_hash_final_spearman_n_permutations: 10000
sensitivity_k11_excluding_hash_final_spearman_permutation_seed: 423424
sensitivity_k11_excluding_hash_final_spearman_bca_95_lower_bound: 0.77496011782487706
sensitivity_k11_excluding_hash_final_spearman_bca_95_upper_bound: 1
sensitivity_k11_excluding_hash_final_spearman_bca_valid_draws: 10000
sensitivity_k11_excluding_hash_final_spearman_bca_requested_draws: 10000
sensitivity_k11_excluding_hash_final_spearman_bca_bootstrap_seed: 423426
sensitivity_k11_excluding_hash_final_spearman_bca_acceleration_defined: 1
secondary_composite_spearman_defined: 1
secondary_composite_spearman_rho: 0.97902097902097907
secondary_composite_spearman_permutation_p_value_two_sided: 9.9990000999900015e-05
secondary_composite_spearman_permutation_extreme_count: 0
secondary_composite_spearman_n_permutations: 10000
secondary_composite_spearman_permutation_seed: 423424
secondary_composite_spearman_bca_95_lower_bound: 0.92619926199261993
secondary_composite_spearman_bca_95_upper_bound: 1
secondary_composite_spearman_bca_valid_draws: 10000
secondary_composite_spearman_bca_requested_draws: 10000
secondary_composite_spearman_bca_bootstrap_seed: 423427
secondary_composite_spearman_bca_acceleration_defined: 1
secondary_nonzero_variance_fraction_spearman_defined: 1
secondary_nonzero_variance_fraction_spearman_rho: -0.41006100983438476
secondary_nonzero_variance_fraction_spearman_permutation_p_value_two_sided: 0.18558144185581441
secondary_nonzero_variance_fraction_spearman_permutation_extreme_count: 1855
secondary_nonzero_variance_fraction_spearman_n_permutations: 10000
secondary_nonzero_variance_fraction_spearman_permutation_seed: 423424
secondary_nonzero_variance_fraction_spearman_bca_95_lower_bound: -0.9310241807718902
secondary_nonzero_variance_fraction_spearman_bca_95_upper_bound: 0.49461221746584805
secondary_nonzero_variance_fraction_spearman_bca_valid_draws: 10000
secondary_nonzero_variance_fraction_spearman_bca_requested_draws: 10000
secondary_nonzero_variance_fraction_spearman_bca_bootstrap_seed: 423428
secondary_nonzero_variance_fraction_spearman_bca_acceleration_defined: 1
movement_outcome_vector_available: 1
movement_outcome_vector_constant: 0
movement_outcome_vector_all_zero: 0
movement_outcome_nonzero_seam_count: 12
movement_outcome_distinct_nonzero_value_count: 10
movement_one_mover_case: 0
movement_degenerate_case_detected: 0
movement_exact_two_sided_p_floor: nan
movement_alpha_0_05_reachable: nan
movement_spearman_defined: 1
movement_spearman_rho: 0.12788652500185485
movement_spearman_permutation_p_value_two_sided: 0.68853114688531147
movement_spearman_permutation_extreme_count: 6885
movement_spearman_n_permutations: 10000
movement_spearman_permutation_seed: 423424
movement_spearman_bca_95_lower_bound: -0.62942740069467573
movement_spearman_bca_95_upper_bound: 0.68199769516885977
movement_spearman_bca_valid_draws: 10000
movement_spearman_bca_requested_draws: 10000
movement_spearman_bca_bootstrap_seed: 423429
movement_spearman_bca_acceleration_defined: 1
movement_spearman_permutation_p_value_floor_aware: 0.68853114688531147
movement_trap_level_rho_threshold: 0.59999999999999998
movement_trap_evaluable: 1
movement_trap_level_correlation_high: 1
movement_trap_no_seam_interval_excludes_zero: 1
movement_trap_no_positive_seam_interval_excludes_zero: 1
movement_trap_high_level_correlation_while_no_seam_moved: 1
positive_control_available: 1
positive_control_screen_emission_rate: 0.61250000000000004
positive_control_initial_20_selection_rate: 0.59027777777777779
positive_control_terminal_20_selection_rate: 0.58194444444444449
positive_control_preregistered_terminal_20_minus_screen: -0.030555555555555558
positive_control_preregistered_cross_instrument_95_lower_bound: -0.08651251584364561
positive_control_preregistered_cross_instrument_95_upper_bound: 0.026222935273766457
positive_control_preregistered_cross_instrument_passed_0_10: 0
positive_control_within_smoke_terminal_20_minus_initial_20: -0.0083333333333333367
positive_control_within_smoke_paired_95_lower_bound: -0.061111111111111116
positive_control_within_smoke_paired_95_upper_bound: 0.048611111111111119
positive_control_within_smoke_passed_0_10: 0
positive_control_criterion_verdicts_agree: 1
positive_control_passed_0_10: 0
binary_rule_ready: 1
binary_classification_correct: 9
binary_classification_total: 12
binary_classification_accuracy: 0.75
binary_classification_wilson_95_lower_bound: 0.46769466506643426
binary_classification_wilson_95_upper_bound: 0.91105833160594529
binary_classification_exact_binomial_p_value_one_sided: 0.072998046875
binary_classification_p_value_holm_adjusted: 0.072998046875
binary_true_positive: 3
binary_true_negative: 6
binary_false_positive: 3
binary_false_negative: 0
binary_misclassified_count: 3
primary_spearman_permutation_p_value_holm_adjusted: 0.00039996000399960006
binary_misclassified_answer_sentence: 1
binary_misclassified_final_answer_prefix: 1
binary_misclassified_answer_prefix: 1
negative_control_available: 1
negative_control_terminal_mean: 0.010185185185185184
negative_control_terminal_minus_gate: -0.23981481481481481
negative_control_difference_95_lower_bound: -0.24166666666666667
negative_control_difference_95_upper_bound: -0.23749999999999999
negative_control_stayed_failed: 1
negative_control_all_seeds_failed: 1
negative_control_seeds_passing_gate: 0
negative_control_majority_label_passed: 0
harness_control_available: 1
harness_control_excluded_from_primary_and_correlations: 1
harness_control_steps: 100
harness_control_initial_20_selection_rate: 0.59166666666666667
harness_control_final_20_selection_rate: 0.60555555555555551
harness_control_selection_success_difference: 10
harness_control_movement_final20_minus_initial20: 0.013888888888888873
harness_control_movement_paired_95_lower_bound: -0.019444444444444466
harness_control_movement_paired_95_upper_bound: 0.047222222222222207
harness_control_movement_interval_excludes_zero: 0
harness_control_max_kl: 0.00047689273681802052
harness_control_kl_spike: 0
harness_control_initial_20_objective_reward_rate: 0.59166666666666667
harness_control_final_20_objective_reward_rate: 0.60555555555555551
harness_control_reward_collapse: 0
harness_control_initial_20_noncapped_completion_rate: 0.94861111111111107
harness_control_final_20_noncapped_completion_rate: 0.95694444444444449
harness_control_termination_collapse: 0
harness_control_destabilized: 0
harness_control_supports_seams_not_reinforceable: 0
harness_control_supports_harness_inert: 1
harness_control_uninformative: 0
harness_control_resume_200_required: 0
harness_control_verdict: supports_harness_inert
harness_control_resume_200_command: set -e; export HF_HOME=~/.cache/huggingface; export CUDA_VISIBLE_DEVICES=0; export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True; ~/projects/incub-b423a72c1602/.venv/bin/python code/run_smoke_leg.py --seam-id boxed_final --seed 0 --steps 200 --reward-mode seam_only --allow-harness-control-resume --output-root results/real/harness_control
leg_health_completed_legs: 18
leg_health_valid_legs: 18
leg_health_suspect_legs: 0
leg_health_flat_but_healthy_legs: 18
leg_health_flat_and_starved_legs: 0
leg_health_answer_prefix_seed0_step_count: 100
leg_health_answer_prefix_seed0_recorded_gpu_hours: 0.45550566828166161
leg_health_answer_prefix_seed0_frac_reward_zero_std_mean: 0.15833333730697632
leg_health_answer_prefix_seed0_frac_reward_zero_std_max: 0.5
leg_health_answer_prefix_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_answer_prefix_seed0_clipped_ratio_mean: 0.048333334233611823
leg_health_answer_prefix_seed0_clipped_ratio_max: 0.27777779102325439
leg_health_answer_prefix_seed0_natural_termination_rate: 0.95166666576638814
leg_health_answer_prefix_seed0_kl_mean: 0.00033218293712707238
leg_health_answer_prefix_seed0_kl_terminal: 0.00032586853679579991
leg_health_answer_prefix_seed0_kl_max: 0.00048415873122091096
leg_health_answer_prefix_seed0_grad_norm_mean: 0.065611743740737444
leg_health_answer_prefix_seed0_grad_norm_median: 0.065227873623371124
leg_health_answer_prefix_seed0_grad_norm_max: 0.088150136172771454
leg_health_answer_prefix_seed0_grad_norm_spike_threshold: 1
leg_health_answer_prefix_seed0_selection_first_20_mean: 0.04027777777777778
leg_health_answer_prefix_seed0_selection_last_20_mean: 0.025000000000000001
leg_health_answer_prefix_seed0_selection_window_success_difference: -11
leg_health_answer_prefix_seed0_selection_movement_last20_minus_first20: -0.015277777777777779
leg_health_answer_prefix_seed0_selection_max_step: 18
leg_health_answer_prefix_seed0_trainer_reward_mean_over_steps: 0.34333333700895308
leg_health_answer_prefix_seed0_trainer_reward_std_over_steps: 0.10202577733561458
leg_health_answer_prefix_seed0_valid: 1
leg_health_answer_prefix_seed0_suspect: 0
leg_health_answer_prefix_seed0_kl_spike: 0
leg_health_answer_prefix_seed0_grad_norm_spike: 0
leg_health_answer_prefix_seed0_variance_healthy: 1
leg_health_answer_prefix_seed0_termination_healthy: 1
leg_health_answer_prefix_seed0_selection_adjacent_monotone: 0
leg_health_answer_prefix_seed0_selection_trajectory_flat: 1
leg_health_answer_prefix_seed0_selection_trajectory_monotone: 0
leg_health_answer_prefix_seed0_selection_trajectory_noisy: 0
leg_health_answer_prefix_seed0_flat_but_healthy: 1
leg_health_answer_prefix_seed0_flat_and_starved: 0
leg_health_answer_sentence_seed0_step_count: 100
leg_health_answer_sentence_seed0_recorded_gpu_hours: 0.45826897453498938
leg_health_answer_sentence_seed0_frac_reward_zero_std_mean: 0.19333333700895308
leg_health_answer_sentence_seed0_frac_reward_zero_std_max: 0.66666668653488159
leg_health_answer_sentence_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_answer_sentence_seed0_clipped_ratio_mean: 0.050833334308117629
leg_health_answer_sentence_seed0_clipped_ratio_max: 0.25
leg_health_answer_sentence_seed0_natural_termination_rate: 0.94916666569188235
leg_health_answer_sentence_seed0_kl_mean: 0.00032732299999527946
leg_health_answer_sentence_seed0_kl_terminal: 0.00031139559966201585
leg_health_answer_sentence_seed0_kl_max: 0.00044476859814797837
leg_health_answer_sentence_seed0_grad_norm_mean: 0.065062046274542804
leg_health_answer_sentence_seed0_grad_norm_median: 0.065010145306587219
leg_health_answer_sentence_seed0_grad_norm_max: 0.0948328897356987
leg_health_answer_sentence_seed0_grad_norm_spike_threshold: 1
leg_health_answer_sentence_seed0_selection_first_20_mean: 0.037499999999999999
leg_health_answer_sentence_seed0_selection_last_20_mean: 0.038888888888888883
leg_health_answer_sentence_seed0_selection_window_success_difference: 1
leg_health_answer_sentence_seed0_selection_movement_last20_minus_first20: 0.0013888888888888887
leg_health_answer_sentence_seed0_selection_max_step: 32
leg_health_answer_sentence_seed0_trainer_reward_mean_over_steps: 0.28944444723427298
leg_health_answer_sentence_seed0_trainer_reward_std_over_steps: 0.10139307502451411
leg_health_answer_sentence_seed0_valid: 1
leg_health_answer_sentence_seed0_suspect: 0
leg_health_answer_sentence_seed0_kl_spike: 0
leg_health_answer_sentence_seed0_grad_norm_spike: 0
leg_health_answer_sentence_seed0_variance_healthy: 1
leg_health_answer_sentence_seed0_termination_healthy: 1
leg_health_answer_sentence_seed0_selection_adjacent_monotone: 0
leg_health_answer_sentence_seed0_selection_trajectory_flat: 1
leg_health_answer_sentence_seed0_selection_trajectory_monotone: 0
leg_health_answer_sentence_seed0_selection_trajectory_noisy: 0
leg_health_answer_sentence_seed0_flat_but_healthy: 1
leg_health_answer_sentence_seed0_flat_and_starved: 0
leg_health_bold_final_seed0_step_count: 100
leg_health_bold_final_seed0_recorded_gpu_hours: 0.46381224120777914
leg_health_bold_final_seed0_frac_reward_zero_std_mean: 0.19000000417232513
leg_health_bold_final_seed0_frac_reward_zero_std_max: 0.83333331346511841
leg_health_bold_final_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_bold_final_seed0_clipped_ratio_mean: 0.051111112069338563
leg_health_bold_final_seed0_clipped_ratio_max: 0.25
leg_health_bold_final_seed0_natural_termination_rate: 0.9488888879306614
leg_health_bold_final_seed0_kl_mean: 0.00031766423625716319
leg_health_bold_final_seed0_kl_terminal: 0.00027532959938980639
leg_health_bold_final_seed0_kl_max: 0.00047191156772896647
leg_health_bold_final_seed0_grad_norm_mean: 0.064880130272358658
leg_health_bold_final_seed0_grad_norm_median: 0.06507149338722229
leg_health_bold_final_seed0_grad_norm_max: 0.084698840975761414
leg_health_bold_final_seed0_grad_norm_spike_threshold: 1
leg_health_bold_final_seed0_selection_first_20_mean: 0.51666666666666672
leg_health_bold_final_seed0_selection_last_20_mean: 0.52222222222222225
leg_health_bold_final_seed0_selection_window_success_difference: 4
leg_health_bold_final_seed0_selection_movement_last20_minus_first20: 0.0055555555555555497
leg_health_bold_final_seed0_selection_max_step: 72
leg_health_bold_final_seed0_trainer_reward_mean_over_steps: 0.69277778267860413
leg_health_bold_final_seed0_trainer_reward_std_over_steps: 0.099522003342279008
leg_health_bold_final_seed0_valid: 1
leg_health_bold_final_seed0_suspect: 0
leg_health_bold_final_seed0_kl_spike: 0
leg_health_bold_final_seed0_grad_norm_spike: 0
leg_health_bold_final_seed0_variance_healthy: 1
leg_health_bold_final_seed0_termination_healthy: 1
leg_health_bold_final_seed0_selection_adjacent_monotone: 0
leg_health_bold_final_seed0_selection_trajectory_flat: 1
leg_health_bold_final_seed0_selection_trajectory_monotone: 0
leg_health_bold_final_seed0_selection_trajectory_noisy: 0
leg_health_bold_final_seed0_flat_but_healthy: 1
leg_health_bold_final_seed0_flat_and_starved: 0
leg_health_boxed_final_seed0_step_count: 100
leg_health_boxed_final_seed0_recorded_gpu_hours: 0.45561788669748543
leg_health_boxed_final_seed0_frac_reward_zero_std_mean: 0.17500000417232514
leg_health_boxed_final_seed0_frac_reward_zero_std_max: 0.5
leg_health_boxed_final_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_boxed_final_seed0_clipped_ratio_mean: 0.039444445129483939
leg_health_boxed_final_seed0_clipped_ratio_max: 0.25
leg_health_boxed_final_seed0_natural_termination_rate: 0.96055555487051603
leg_health_boxed_final_seed0_kl_mean: 0.00032063646557314011
leg_health_boxed_final_seed0_kl_terminal: 0.00031018432734223705
leg_health_boxed_final_seed0_kl_max: 0.00042693766105609637
leg_health_boxed_final_seed0_grad_norm_mean: 0.066017425507307048
leg_health_boxed_final_seed0_grad_norm_median: 0.065606437623500824
leg_health_boxed_final_seed0_grad_norm_max: 0.092733509838581085
leg_health_boxed_final_seed0_grad_norm_spike_threshold: 1
leg_health_boxed_final_seed0_selection_first_20_mean: 0.59027777777777779
leg_health_boxed_final_seed0_selection_last_20_mean: 0.58194444444444449
leg_health_boxed_final_seed0_selection_window_success_difference: -6
leg_health_boxed_final_seed0_selection_movement_last20_minus_first20: -0.0083333333333333367
leg_health_boxed_final_seed0_selection_max_step: 67
leg_health_boxed_final_seed0_trainer_reward_mean_over_steps: 0.69777778416872027
leg_health_boxed_final_seed0_trainer_reward_std_over_steps: 0.082240989149511445
leg_health_boxed_final_seed0_valid: 1
leg_health_boxed_final_seed0_suspect: 0
leg_health_boxed_final_seed0_kl_spike: 0
leg_health_boxed_final_seed0_grad_norm_spike: 0
leg_health_boxed_final_seed0_variance_healthy: 1
leg_health_boxed_final_seed0_termination_healthy: 1
leg_health_boxed_final_seed0_selection_adjacent_monotone: 0
leg_health_boxed_final_seed0_selection_trajectory_flat: 1
leg_health_boxed_final_seed0_selection_trajectory_monotone: 0
leg_health_boxed_final_seed0_selection_trajectory_noisy: 0
leg_health_boxed_final_seed0_flat_but_healthy: 1
leg_health_boxed_final_seed0_flat_and_starved: 0
leg_health_caps_verified_seed0_step_count: 100
leg_health_caps_verified_seed0_recorded_gpu_hours: 0.45680343411416591
leg_health_caps_verified_seed0_frac_reward_zero_std_mean: 0.19000000432133674
leg_health_caps_verified_seed0_frac_reward_zero_std_max: 0.66666668653488159
leg_health_caps_verified_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_caps_verified_seed0_clipped_ratio_mean: 0.052500000912696125
leg_health_caps_verified_seed0_clipped_ratio_max: 0.3055555522441864
leg_health_caps_verified_seed0_natural_termination_rate: 0.94749999908730387
leg_health_caps_verified_seed0_kl_mean: 0.00032299747205494591
leg_health_caps_verified_seed0_kl_terminal: 0.00029066930680225295
leg_health_caps_verified_seed0_kl_max: 0.00043474492849782109
leg_health_caps_verified_seed0_grad_norm_mean: 0.064588041007518762
leg_health_caps_verified_seed0_grad_norm_median: 0.065002921968698502
leg_health_caps_verified_seed0_grad_norm_max: 0.095226861536502838
leg_health_caps_verified_seed0_grad_norm_spike_threshold: 1
leg_health_caps_verified_seed0_selection_first_20_mean: 0.02361111111111111
leg_health_caps_verified_seed0_selection_last_20_mean: 0.016666666666666666
leg_health_caps_verified_seed0_selection_window_success_difference: -5
leg_health_caps_verified_seed0_selection_movement_last20_minus_first20: -0.0069444444444444449
leg_health_caps_verified_seed0_selection_max_step: 7
leg_health_caps_verified_seed0_trainer_reward_mean_over_steps: 0.29638889290392401
leg_health_caps_verified_seed0_trainer_reward_std_over_steps: 0.090442584186481448
leg_health_caps_verified_seed0_valid: 1
leg_health_caps_verified_seed0_suspect: 0
leg_health_caps_verified_seed0_kl_spike: 0
leg_health_caps_verified_seed0_grad_norm_spike: 0
leg_health_caps_verified_seed0_variance_healthy: 1
leg_health_caps_verified_seed0_termination_healthy: 1
leg_health_caps_verified_seed0_selection_adjacent_monotone: 0
leg_health_caps_verified_seed0_selection_trajectory_flat: 1
leg_health_caps_verified_seed0_selection_trajectory_monotone: 0
leg_health_caps_verified_seed0_selection_trajectory_noisy: 0
leg_health_caps_verified_seed0_flat_but_healthy: 1
leg_health_caps_verified_seed0_flat_and_starved: 0
leg_health_caps_verified_seed1_step_count: 100
leg_health_caps_verified_seed1_recorded_gpu_hours: 0.45574328452778801
leg_health_caps_verified_seed1_frac_reward_zero_std_mean: 0.17500000447034836
leg_health_caps_verified_seed1_frac_reward_zero_std_max: 0.66666668653488159
leg_health_caps_verified_seed1_all_groups_zero_variance_step_fraction: 0
leg_health_caps_verified_seed1_clipped_ratio_mean: 0.050000000949949025
leg_health_caps_verified_seed1_clipped_ratio_max: 0.2222222238779068
leg_health_caps_verified_seed1_natural_termination_rate: 0.94999999905005095
leg_health_caps_verified_seed1_kl_mean: 0.00032397631541243756
leg_health_caps_verified_seed1_kl_terminal: 0.00032583593080441159
leg_health_caps_verified_seed1_kl_max: 0.00048935688876857364
leg_health_caps_verified_seed1_grad_norm_mean: 0.065102320164442062
leg_health_caps_verified_seed1_grad_norm_median: 0.065964695066213608
leg_health_caps_verified_seed1_grad_norm_max: 0.093851521611213684
leg_health_caps_verified_seed1_grad_norm_spike_threshold: 1
leg_health_caps_verified_seed1_selection_first_20_mean: 0.030555555555555551
leg_health_caps_verified_seed1_selection_last_20_mean: 0.020833333333333332
leg_health_caps_verified_seed1_selection_window_success_difference: -7
leg_health_caps_verified_seed1_selection_movement_last20_minus_first20: -0.0097222222222222206
leg_health_caps_verified_seed1_selection_max_step: 4
leg_health_caps_verified_seed1_trainer_reward_mean_over_steps: 0.3097222254425287
leg_health_caps_verified_seed1_trainer_reward_std_over_steps: 0.10117938747734428
leg_health_caps_verified_seed1_valid: 1
leg_health_caps_verified_seed1_suspect: 0
leg_health_caps_verified_seed1_kl_spike: 0
leg_health_caps_verified_seed1_grad_norm_spike: 0
leg_health_caps_verified_seed1_variance_healthy: 1
leg_health_caps_verified_seed1_termination_healthy: 1
leg_health_caps_verified_seed1_selection_adjacent_monotone: 0
leg_health_caps_verified_seed1_selection_trajectory_flat: 1
leg_health_caps_verified_seed1_selection_trajectory_monotone: 0
leg_health_caps_verified_seed1_selection_trajectory_noisy: 0
leg_health_caps_verified_seed1_flat_but_healthy: 1
leg_health_caps_verified_seed1_flat_and_starved: 0
leg_health_caps_verified_seed2_step_count: 100
leg_health_caps_verified_seed2_recorded_gpu_hours: 0.46188970395305662
leg_health_caps_verified_seed2_frac_reward_zero_std_mean: 0.19666667193174361
leg_health_caps_verified_seed2_frac_reward_zero_std_max: 0.66666668653488159
leg_health_caps_verified_seed2_all_groups_zero_variance_step_fraction: 0
leg_health_caps_verified_seed2_clipped_ratio_mean: 0.050277778767049315
leg_health_caps_verified_seed2_clipped_ratio_max: 0.2222222238779068
leg_health_caps_verified_seed2_natural_termination_rate: 0.94972222123295069
leg_health_caps_verified_seed2_kl_mean: 0.00033051573618043527
leg_health_caps_verified_seed2_kl_terminal: 0.00041994527176332969
leg_health_caps_verified_seed2_kl_max: 0.00047743462103729445
leg_health_caps_verified_seed2_grad_norm_mean: 0.066934371069073684
leg_health_caps_verified_seed2_grad_norm_median: 0.068774387240409851
leg_health_caps_verified_seed2_grad_norm_max: 0.091822363436222076
leg_health_caps_verified_seed2_grad_norm_spike_threshold: 1
leg_health_caps_verified_seed2_selection_first_20_mean: 0.015277777777777776
leg_health_caps_verified_seed2_selection_last_20_mean: 0.025000000000000001
leg_health_caps_verified_seed2_selection_window_success_difference: 7
leg_health_caps_verified_seed2_selection_movement_last20_minus_first20: 0.0097222222222222206
leg_health_caps_verified_seed2_selection_max_step: 38
leg_health_caps_verified_seed2_trainer_reward_mean_over_steps: 0.29638889100402593
leg_health_caps_verified_seed2_trainer_reward_std_over_steps: 0.097266259413291728
leg_health_caps_verified_seed2_valid: 1
leg_health_caps_verified_seed2_suspect: 0
leg_health_caps_verified_seed2_kl_spike: 0
leg_health_caps_verified_seed2_grad_norm_spike: 0
leg_health_caps_verified_seed2_variance_healthy: 1
leg_health_caps_verified_seed2_termination_healthy: 1
leg_health_caps_verified_seed2_selection_adjacent_monotone: 0
leg_health_caps_verified_seed2_selection_trajectory_flat: 1
leg_health_caps_verified_seed2_selection_trajectory_monotone: 0
leg_health_caps_verified_seed2_selection_trajectory_noisy: 0
leg_health_caps_verified_seed2_flat_but_healthy: 1
leg_health_caps_verified_seed2_flat_and_starved: 0
leg_health_final_answer_prefix_seed0_step_count: 100
leg_health_final_answer_prefix_seed0_recorded_gpu_hours: 0.4584232909019435
leg_health_final_answer_prefix_seed0_frac_reward_zero_std_mean: 0.17000000372529031
leg_health_final_answer_prefix_seed0_frac_reward_zero_std_max: 0.66666668653488159
leg_health_final_answer_prefix_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_final_answer_prefix_seed0_clipped_ratio_mean: 0.045277778562158345
leg_health_final_answer_prefix_seed0_clipped_ratio_max: 0.25
leg_health_final_answer_prefix_seed0_natural_termination_rate: 0.95472222143784169
leg_health_final_answer_prefix_seed0_kl_mean: 0.00032525994388076163
leg_health_final_answer_prefix_seed0_kl_terminal: 0.00035610095558998484
leg_health_final_answer_prefix_seed0_kl_max: 0.00044156252018486458
leg_health_final_answer_prefix_seed0_grad_norm_mean: 0.0651371419429779
leg_health_final_answer_prefix_seed0_grad_norm_median: 0.067303631454706192
leg_health_final_answer_prefix_seed0_grad_norm_max: 0.088411778211593628
leg_health_final_answer_prefix_seed0_grad_norm_spike_threshold: 1
leg_health_final_answer_prefix_seed0_selection_first_20_mean: 0.026388888888888889
leg_health_final_answer_prefix_seed0_selection_last_20_mean: 0.041666666666666664
leg_health_final_answer_prefix_seed0_selection_window_success_difference: 11
leg_health_final_answer_prefix_seed0_selection_movement_last20_minus_first20: 0.015277777777777776
leg_health_final_answer_prefix_seed0_selection_max_step: 91
leg_health_final_answer_prefix_seed0_trainer_reward_mean_over_steps: 0.32222222402691841
leg_health_final_answer_prefix_seed0_trainer_reward_std_over_steps: 0.097341195426331989
leg_health_final_answer_prefix_seed0_valid: 1
leg_health_final_answer_prefix_seed0_suspect: 0
leg_health_final_answer_prefix_seed0_kl_spike: 0
leg_health_final_answer_prefix_seed0_grad_norm_spike: 0
leg_health_final_answer_prefix_seed0_variance_healthy: 1
leg_health_final_answer_prefix_seed0_termination_healthy: 1
leg_health_final_answer_prefix_seed0_selection_adjacent_monotone: 0
leg_health_final_answer_prefix_seed0_selection_trajectory_flat: 1
leg_health_final_answer_prefix_seed0_selection_trajectory_monotone: 0
leg_health_final_answer_prefix_seed0_selection_trajectory_noisy: 0
leg_health_final_answer_prefix_seed0_flat_but_healthy: 1
leg_health_final_answer_prefix_seed0_flat_and_starved: 0
leg_health_final_tag_seed0_step_count: 100
leg_health_final_tag_seed0_recorded_gpu_hours: 0.45479936118944253
leg_health_final_tag_seed0_frac_reward_zero_std_mean: 0.14500000417232514
leg_health_final_tag_seed0_frac_reward_zero_std_max: 0.5
leg_health_final_tag_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_final_tag_seed0_clipped_ratio_mean: 0.049722223151475189
leg_health_final_tag_seed0_clipped_ratio_max: 0.25
leg_health_final_tag_seed0_natural_termination_rate: 0.95027777684852477
leg_health_final_tag_seed0_kl_mean: 0.0003246196898786972
leg_health_final_tag_seed0_kl_terminal: 0.00030131964012980461
leg_health_final_tag_seed0_kl_max: 0.00043340748137173552
leg_health_final_tag_seed0_grad_norm_mean: 0.065196558348834519
leg_health_final_tag_seed0_grad_norm_median: 0.065037775784730911
leg_health_final_tag_seed0_grad_norm_max: 0.094404786825180054
leg_health_final_tag_seed0_grad_norm_spike_threshold: 1
leg_health_final_tag_seed0_selection_first_20_mean: 0.01111111111111111
leg_health_final_tag_seed0_selection_last_20_mean: 0.01388888888888889
leg_health_final_tag_seed0_selection_window_success_difference: 2
leg_health_final_tag_seed0_selection_movement_last20_minus_first20: 0.0027777777777777775
leg_health_final_tag_seed0_selection_max_step: 48
leg_health_final_tag_seed0_trainer_reward_mean_over_steps: 0.44166667103767393
leg_health_final_tag_seed0_trainer_reward_std_over_steps: 0.10352580697921754
leg_health_final_tag_seed0_valid: 1
leg_health_final_tag_seed0_suspect: 0
leg_health_final_tag_seed0_kl_spike: 0
leg_health_final_tag_seed0_grad_norm_spike: 0
leg_health_final_tag_seed0_variance_healthy: 1
leg_health_final_tag_seed0_termination_healthy: 1
leg_health_final_tag_seed0_selection_adjacent_monotone: 0
leg_health_final_tag_seed0_selection_trajectory_flat: 1
leg_health_final_tag_seed0_selection_trajectory_monotone: 0
leg_health_final_tag_seed0_selection_trajectory_noisy: 0
leg_health_final_tag_seed0_flat_but_healthy: 1
leg_health_final_tag_seed0_flat_and_starved: 0
leg_health_hash_final_seed0_step_count: 100
leg_health_hash_final_seed0_recorded_gpu_hours: 0.45692385212972619
leg_health_hash_final_seed0_frac_reward_zero_std_mean: 0.078333335667848586
leg_health_hash_final_seed0_frac_reward_zero_std_max: 0.3333333432674408
leg_health_hash_final_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_hash_final_seed0_clipped_ratio_mean: 0.044722223076969388
leg_health_hash_final_seed0_clipped_ratio_max: 0.2222222238779068
leg_health_hash_final_seed0_natural_termination_rate: 0.95527777692303062
leg_health_hash_final_seed0_kl_mean: 0.00032336736340463779
leg_health_hash_final_seed0_kl_terminal: 0.00031153434732307989
leg_health_hash_final_seed0_kl_max: 0.0004377984635842343
leg_health_hash_final_seed0_grad_norm_mean: 0.071367466859519488
leg_health_hash_final_seed0_grad_norm_median: 0.070996969938278198
leg_health_hash_final_seed0_grad_norm_max: 0.097159333527088165
leg_health_hash_final_seed0_grad_norm_spike_threshold: 1
leg_health_hash_final_seed0_selection_first_20_mean: 0.56944444444444442
leg_health_hash_final_seed0_selection_last_20_mean: 0.58750000000000002
leg_health_hash_final_seed0_selection_window_success_difference: 13
leg_health_hash_final_seed0_selection_movement_last20_minus_first20: 0.018055555555555554
leg_health_hash_final_seed0_selection_max_step: 18
leg_health_hash_final_seed0_trainer_reward_mean_over_steps: 0.62555556267499923
leg_health_hash_final_seed0_trainer_reward_std_over_steps: 0.10019117852736711
leg_health_hash_final_seed0_valid: 1
leg_health_hash_final_seed0_suspect: 0
leg_health_hash_final_seed0_kl_spike: 0
leg_health_hash_final_seed0_grad_norm_spike: 0
leg_health_hash_final_seed0_variance_healthy: 1
leg_health_hash_final_seed0_termination_healthy: 1
leg_health_hash_final_seed0_selection_adjacent_monotone: 0
leg_health_hash_final_seed0_selection_trajectory_flat: 1
leg_health_hash_final_seed0_selection_trajectory_monotone: 0
leg_health_hash_final_seed0_selection_trajectory_noisy: 0
leg_health_hash_final_seed0_flat_but_healthy: 1
leg_health_hash_final_seed0_flat_and_starved: 0
leg_health_hex_sentinel_seed0_step_count: 100
leg_health_hex_sentinel_seed0_recorded_gpu_hours: 0.46682085837611037
leg_health_hex_sentinel_seed0_frac_reward_zero_std_mean: 0.14333333700895309
leg_health_hex_sentinel_seed0_frac_reward_zero_std_max: 0.66666668653488159
leg_health_hex_sentinel_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_hex_sentinel_seed0_clipped_ratio_mean: 0.054166667815297845
leg_health_hex_sentinel_seed0_clipped_ratio_max: 0.27777779102325439
leg_health_hex_sentinel_seed0_natural_termination_rate: 0.94583333218470211
leg_health_hex_sentinel_seed0_kl_mean: 0.00033445593643894732
leg_health_hex_sentinel_seed0_kl_terminal: 0.00032274827632742625
leg_health_hex_sentinel_seed0_kl_max: 0.00044416789508735138
leg_health_hex_sentinel_seed0_grad_norm_mean: 0.063928033132106071
leg_health_hex_sentinel_seed0_grad_norm_median: 0.065229602158069611
leg_health_hex_sentinel_seed0_grad_norm_max: 0.091584205627441406
leg_health_hex_sentinel_seed0_grad_norm_spike_threshold: 1
leg_health_hex_sentinel_seed0_selection_first_20_mean: 0.0027777777777777775
leg_health_hex_sentinel_seed0_selection_last_20_mean: 0
leg_health_hex_sentinel_seed0_selection_window_success_difference: -2
leg_health_hex_sentinel_seed0_selection_movement_last20_minus_first20: -0.0027777777777777775
leg_health_hex_sentinel_seed0_selection_max_step: 14
leg_health_hex_sentinel_seed0_trainer_reward_mean_over_steps: 0.40250000357627869
leg_health_hex_sentinel_seed0_trainer_reward_std_over_steps: 0.11176854138629785
leg_health_hex_sentinel_seed0_valid: 1
leg_health_hex_sentinel_seed0_suspect: 0
leg_health_hex_sentinel_seed0_kl_spike: 0
leg_health_hex_sentinel_seed0_grad_norm_spike: 0
leg_health_hex_sentinel_seed0_variance_healthy: 1
leg_health_hex_sentinel_seed0_termination_healthy: 1
leg_health_hex_sentinel_seed0_selection_adjacent_monotone: 0
leg_health_hex_sentinel_seed0_selection_trajectory_flat: 1
leg_health_hex_sentinel_seed0_selection_trajectory_monotone: 0
leg_health_hex_sentinel_seed0_selection_trajectory_noisy: 0
leg_health_hex_sentinel_seed0_flat_but_healthy: 1
leg_health_hex_sentinel_seed0_flat_and_starved: 0
leg_health_numeric_only_seed0_step_count: 100
leg_health_numeric_only_seed0_recorded_gpu_hours: 0.45502934300111114
leg_health_numeric_only_seed0_frac_reward_zero_std_mean: 0.14833333671092988
leg_health_numeric_only_seed0_frac_reward_zero_std_max: 0.66666668653488159
leg_health_numeric_only_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_numeric_only_seed0_clipped_ratio_mean: 0.047777778785675767
leg_health_numeric_only_seed0_clipped_ratio_max: 0.2222222238779068
leg_health_numeric_only_seed0_natural_termination_rate: 0.9522222212143242
leg_health_numeric_only_seed0_kl_mean: 0.00031782668971573001
leg_health_numeric_only_seed0_kl_terminal: 0.00031526303791906685
leg_health_numeric_only_seed0_kl_max: 0.00043186421195665997
leg_health_numeric_only_seed0_grad_norm_mean: 0.064955536872148514
leg_health_numeric_only_seed0_grad_norm_median: 0.064313136041164398
leg_health_numeric_only_seed0_grad_norm_max: 0.095608554780483246
leg_health_numeric_only_seed0_grad_norm_spike_threshold: 1
leg_health_numeric_only_seed0_selection_first_20_mean: 0.0027777777777777775
leg_health_numeric_only_seed0_selection_last_20_mean: 0.0055555555555555549
leg_health_numeric_only_seed0_selection_window_success_difference: 2
leg_health_numeric_only_seed0_selection_movement_last20_minus_first20: 0.0027777777777777775
leg_health_numeric_only_seed0_selection_max_step: 39
leg_health_numeric_only_seed0_trainer_reward_mean_over_steps: 0.45277778297662735
leg_health_numeric_only_seed0_trainer_reward_std_over_steps: 0.11716165091033506
leg_health_numeric_only_seed0_valid: 1
leg_health_numeric_only_seed0_suspect: 0
leg_health_numeric_only_seed0_kl_spike: 0
leg_health_numeric_only_seed0_grad_norm_spike: 0
leg_health_numeric_only_seed0_variance_healthy: 1
leg_health_numeric_only_seed0_termination_healthy: 1
leg_health_numeric_only_seed0_selection_adjacent_monotone: 0
leg_health_numeric_only_seed0_selection_trajectory_flat: 1
leg_health_numeric_only_seed0_selection_trajectory_monotone: 0
leg_health_numeric_only_seed0_selection_trajectory_noisy: 0
leg_health_numeric_only_seed0_flat_but_healthy: 1
leg_health_numeric_only_seed0_flat_and_starved: 0
leg_health_parent_directive_seed0_step_count: 100
leg_health_parent_directive_seed0_recorded_gpu_hours: 0.46345581150888593
leg_health_parent_directive_seed0_frac_reward_zero_std_mean: 0.22166667163372039
leg_health_parent_directive_seed0_frac_reward_zero_std_max: 0.5
leg_health_parent_directive_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_parent_directive_seed0_clipped_ratio_mean: 0.052222223207354548
leg_health_parent_directive_seed0_clipped_ratio_max: 0.25
leg_health_parent_directive_seed0_natural_termination_rate: 0.94777777679264541
leg_health_parent_directive_seed0_kl_mean: 0.00032873871935104636
leg_health_parent_directive_seed0_kl_terminal: 0.0003336655693904807
leg_health_parent_directive_seed0_kl_max: 0.00045264425959127647
leg_health_parent_directive_seed0_grad_norm_mean: 0.062494673840701578
leg_health_parent_directive_seed0_grad_norm_median: 0.062025617808103561
leg_health_parent_directive_seed0_grad_norm_max: 0.088312007486820221
leg_health_parent_directive_seed0_grad_norm_spike_threshold: 1
leg_health_parent_directive_seed0_selection_first_20_mean: 0.0097222222222222206
leg_health_parent_directive_seed0_selection_last_20_mean: 0.0083333333333333332
leg_health_parent_directive_seed0_selection_window_success_difference: -1
leg_health_parent_directive_seed0_selection_movement_last20_minus_first20: -0.0013888888888888887
leg_health_parent_directive_seed0_selection_max_step: 90
leg_health_parent_directive_seed0_trainer_reward_mean_over_steps: 0.26944444864988326
leg_health_parent_directive_seed0_trainer_reward_std_over_steps: 0.084391430347181481
leg_health_parent_directive_seed0_valid: 1
leg_health_parent_directive_seed0_suspect: 0
leg_health_parent_directive_seed0_kl_spike: 0
leg_health_parent_directive_seed0_grad_norm_spike: 0
leg_health_parent_directive_seed0_variance_healthy: 1
leg_health_parent_directive_seed0_termination_healthy: 1
leg_health_parent_directive_seed0_selection_adjacent_monotone: 0
leg_health_parent_directive_seed0_selection_trajectory_flat: 1
leg_health_parent_directive_seed0_selection_trajectory_monotone: 0
leg_health_parent_directive_seed0_selection_trajectory_noisy: 0
leg_health_parent_directive_seed0_flat_but_healthy: 1
leg_health_parent_directive_seed0_flat_and_starved: 0
leg_health_parent_directive_seed1_step_count: 100
leg_health_parent_directive_seed1_recorded_gpu_hours: 0.46597963777888152
leg_health_parent_directive_seed1_frac_reward_zero_std_mean: 0.24166667133569716
leg_health_parent_directive_seed1_frac_reward_zero_std_max: 0.83333331346511841
leg_health_parent_directive_seed1_all_groups_zero_variance_step_fraction: 0
leg_health_parent_directive_seed1_clipped_ratio_mean: 0.058333334624767301
leg_health_parent_directive_seed1_clipped_ratio_max: 0.2222222238779068
leg_health_parent_directive_seed1_natural_termination_rate: 0.94166666537523269
leg_health_parent_directive_seed1_kl_mean: 0.00032731747308086293
leg_health_parent_directive_seed1_kl_terminal: 0.00033224241148370009
leg_health_parent_directive_seed1_kl_max: 0.000427029929899921
leg_health_parent_directive_seed1_grad_norm_mean: 0.061640582364052537
leg_health_parent_directive_seed1_grad_norm_median: 0.062215154990553856
leg_health_parent_directive_seed1_grad_norm_max: 0.095287531614303589
leg_health_parent_directive_seed1_grad_norm_spike_threshold: 1
leg_health_parent_directive_seed1_selection_first_20_mean: 0.012500000000000001
leg_health_parent_directive_seed1_selection_last_20_mean: 0.012500000000000001
leg_health_parent_directive_seed1_selection_window_success_difference: 0
leg_health_parent_directive_seed1_selection_movement_last20_minus_first20: 0
leg_health_parent_directive_seed1_selection_max_step: 3
leg_health_parent_directive_seed1_trainer_reward_mean_over_steps: 0.26583333633840084
leg_health_parent_directive_seed1_trainer_reward_std_over_steps: 0.089516447436779154
leg_health_parent_directive_seed1_valid: 1
leg_health_parent_directive_seed1_suspect: 0
leg_health_parent_directive_seed1_kl_spike: 0
leg_health_parent_directive_seed1_grad_norm_spike: 0
leg_health_parent_directive_seed1_variance_healthy: 1
leg_health_parent_directive_seed1_termination_healthy: 1
leg_health_parent_directive_seed1_selection_adjacent_monotone: 0
leg_health_parent_directive_seed1_selection_trajectory_flat: 1
leg_health_parent_directive_seed1_selection_trajectory_monotone: 0
leg_health_parent_directive_seed1_selection_trajectory_noisy: 0
leg_health_parent_directive_seed1_flat_but_healthy: 1
leg_health_parent_directive_seed1_flat_and_starved: 0
leg_health_parent_directive_seed2_step_count: 100
leg_health_parent_directive_seed2_recorded_gpu_hours: 0.4604265013699963
leg_health_parent_directive_seed2_frac_reward_zero_std_mean: 0.24166667178273202
leg_health_parent_directive_seed2_frac_reward_zero_std_max: 0.66666668653488159
leg_health_parent_directive_seed2_all_groups_zero_variance_step_fraction: 0
leg_health_parent_directive_seed2_clipped_ratio_mean: 0.053333334363996981
leg_health_parent_directive_seed2_clipped_ratio_max: 0.2222222238779068
leg_health_parent_directive_seed2_natural_termination_rate: 0.94666666563600299
leg_health_parent_directive_seed2_kl_mean: 0.00033676839453012996
leg_health_parent_directive_seed2_kl_terminal: 0.00040209435974247754
leg_health_parent_directive_seed2_kl_max: 0.0004675301315728575
leg_health_parent_directive_seed2_grad_norm_mean: 0.063619510047137742
leg_health_parent_directive_seed2_grad_norm_median: 0.062644191086292267
leg_health_parent_directive_seed2_grad_norm_max: 0.092143237590789795
leg_health_parent_directive_seed2_grad_norm_spike_threshold: 1
leg_health_parent_directive_seed2_selection_first_20_mean: 0.0083333333333333332
leg_health_parent_directive_seed2_selection_last_20_mean: 0.0097222222222222206
leg_health_parent_directive_seed2_selection_window_success_difference: 1
leg_health_parent_directive_seed2_selection_movement_last20_minus_first20: 0.0013888888888888887
leg_health_parent_directive_seed2_selection_max_step: 33
leg_health_parent_directive_seed2_trainer_reward_mean_over_steps: 0.25388889204710724
leg_health_parent_directive_seed2_trainer_reward_std_over_steps: 0.084985474384864354
leg_health_parent_directive_seed2_valid: 1
leg_health_parent_directive_seed2_suspect: 0
leg_health_parent_directive_seed2_kl_spike: 0
leg_health_parent_directive_seed2_grad_norm_spike: 0
leg_health_parent_directive_seed2_variance_healthy: 1
leg_health_parent_directive_seed2_termination_healthy: 1
leg_health_parent_directive_seed2_selection_adjacent_monotone: 0
leg_health_parent_directive_seed2_selection_trajectory_flat: 1
leg_health_parent_directive_seed2_selection_trajectory_monotone: 0
leg_health_parent_directive_seed2_selection_trajectory_noisy: 0
leg_health_parent_directive_seed2_flat_but_healthy: 1
leg_health_parent_directive_seed2_flat_and_starved: 0
leg_health_result_prefix_seed0_step_count: 100
leg_health_result_prefix_seed0_recorded_gpu_hours: 0.45613804283750747
leg_health_result_prefix_seed0_frac_reward_zero_std_mean: 0.14666667059063912
leg_health_result_prefix_seed0_frac_reward_zero_std_max: 0.66666668653488159
leg_health_result_prefix_seed0_all_groups_zero_variance_step_fraction: 0
leg_health_result_prefix_seed0_clipped_ratio_mean: 0.051111111976206304
leg_health_result_prefix_seed0_clipped_ratio_max: 0.25
leg_health_result_prefix_seed0_natural_termination_rate: 0.94888888802379368
leg_health_result_prefix_seed0_kl_mean: 0.0003249629605852533
leg_health_result_prefix_seed0_kl_terminal: 0.00035498530875581008
leg_health_result_prefix_seed0_kl_max: 0.00045823144804065424
leg_health_result_prefix_seed0_grad_norm_mean: 0.065165869146585467
leg_health_result_prefix_seed0_grad_norm_median: 0.064270235598087311
leg_health_result_prefix_seed0_grad_norm_max: 0.10082994401454926
leg_health_result_prefix_seed0_grad_norm_spike_threshold: 1
leg_health_result_prefix_seed0_selection_first_20_mean: 0.018055555555555554
leg_health_result_prefix_seed0_selection_last_20_mean: 0.020833333333333332
leg_health_result_prefix_seed0_selection_window_success_difference: 2
leg_health_result_prefix_seed0_selection_movement_last20_minus_first20: 0.0027777777777777775
leg_health_result_prefix_seed0_selection_max_step: 2
leg_health_result_prefix_seed0_trainer_reward_mean_over_steps: 0.39361111618578432
leg_health_result_prefix_seed0_trainer_reward_std_over_steps: 0.11686854371608606
leg_health_result_prefix_seed0_valid: 1
leg_health_result_prefix_seed0_suspect: 0
leg_health_result_prefix_seed0_kl_spike: 0
leg_health_result_prefix_seed0_grad_norm_spike: 0
leg_health_result_prefix_seed0_variance_healthy: 1
leg_health_result_prefix_seed0_termination_healthy: 1
leg_health_result_prefix_seed0_selection_adjacent_monotone: 0
leg_health_result_prefix_seed0_selection_trajectory_flat: 1
leg_health_result_prefix_seed0_selection_trajectory_monotone: 0
leg_health_result_prefix_seed0_selection_trajectory_noisy: 0
leg_health_result_prefix_seed0_flat_but_healthy: 1
leg_health_result_prefix_seed0_flat_and_starved: 0
leg_health_result_prefix_seed1_step_count: 100
leg_health_result_prefix_seed1_recorded_gpu_hours: 0.45497691789027561
leg_health_result_prefix_seed1_frac_reward_zero_std_mean: 0.16833333730697631
leg_health_result_prefix_seed1_frac_reward_zero_std_max: 0.5
leg_health_result_prefix_seed1_all_groups_zero_variance_step_fraction: 0
leg_health_result_prefix_seed1_clipped_ratio_mean: 0.055277779120951893
leg_health_result_prefix_seed1_clipped_ratio_max: 0.27777779102325439
leg_health_result_prefix_seed1_natural_termination_rate: 0.94472222087904811
leg_health_result_prefix_seed1_kl_mean: 0.00031667679024394602
leg_health_result_prefix_seed1_kl_terminal: 0.00032044375742164749
leg_health_result_prefix_seed1_kl_max: 0.00042233416267360252
leg_health_result_prefix_seed1_grad_norm_mean: 0.06425725214183331
leg_health_result_prefix_seed1_grad_norm_median: 0.064802899956703186
leg_health_result_prefix_seed1_grad_norm_max: 0.096243567764759064
leg_health_result_prefix_seed1_grad_norm_spike_threshold: 1
leg_health_result_prefix_seed1_selection_first_20_mean: 0.02361111111111111
leg_health_result_prefix_seed1_selection_last_20_mean: 0.015277777777777776
leg_health_result_prefix_seed1_selection_window_success_difference: -6
leg_health_result_prefix_seed1_selection_movement_last20_minus_first20: -0.0083333333333333332
leg_health_result_prefix_seed1_selection_max_step: 7
leg_health_result_prefix_seed1_trainer_reward_mean_over_steps: 0.38722222499549391
leg_health_result_prefix_seed1_trainer_reward_std_over_steps: 0.093856012626879121
leg_health_result_prefix_seed1_valid: 1
leg_health_result_prefix_seed1_suspect: 0
leg_health_result_prefix_seed1_kl_spike: 0
leg_health_result_prefix_seed1_grad_norm_spike: 0
leg_health_result_prefix_seed1_variance_healthy: 1
leg_health_result_prefix_seed1_termination_healthy: 1
leg_health_result_prefix_seed1_selection_adjacent_monotone: 0
leg_health_result_prefix_seed1_selection_trajectory_flat: 1
leg_health_result_prefix_seed1_selection_trajectory_monotone: 0
leg_health_result_prefix_seed1_selection_trajectory_noisy: 0
leg_health_result_prefix_seed1_flat_but_healthy: 1
leg_health_result_prefix_seed1_flat_and_starved: 0
leg_health_result_prefix_seed2_step_count: 100
leg_health_result_prefix_seed2_recorded_gpu_hours: 0.45710208162917598
leg_health_result_prefix_seed2_frac_reward_zero_std_mean: 0.13833333685994148
leg_health_result_prefix_seed2_frac_reward_zero_std_max: 0.5
leg_health_result_prefix_seed2_all_groups_zero_variance_step_fraction: 0
leg_health_result_prefix_seed2_clipped_ratio_mean: 0.052500001024454832
leg_health_result_prefix_seed2_clipped_ratio_max: 0.2222222238779068
leg_health_result_prefix_seed2_natural_termination_rate: 0.94749999897554515
leg_health_result_prefix_seed2_kl_mean: 0.0003246477009573331
leg_health_result_prefix_seed2_kl_terminal: 0.00041023380860375863
leg_health_result_prefix_seed2_kl_max: 0.00045671637053601444
leg_health_result_prefix_seed2_grad_norm_mean: 0.068243779912590979
leg_health_result_prefix_seed2_grad_norm_median: 0.067963868379592896
leg_health_result_prefix_seed2_grad_norm_max: 0.096469566226005554
leg_health_result_prefix_seed2_grad_norm_spike_threshold: 1
leg_health_result_prefix_seed2_selection_first_20_mean: 0.01388888888888889
leg_health_result_prefix_seed2_selection_last_20_mean: 0.01388888888888889
leg_health_result_prefix_seed2_selection_window_success_difference: 0
leg_health_result_prefix_seed2_selection_movement_last20_minus_first20: 0
leg_health_result_prefix_seed2_selection_max_step: 49
leg_health_result_prefix_seed2_trainer_reward_mean_over_steps: 0.39583333685994149
leg_health_result_prefix_seed2_trainer_reward_std_over_steps: 0.090810395785222986
leg_health_result_prefix_seed2_valid: 1
leg_health_result_prefix_seed2_suspect: 0
leg_health_result_prefix_seed2_kl_spike: 0
leg_health_result_prefix_seed2_grad_norm_spike: 0
leg_health_result_prefix_seed2_variance_healthy: 1
leg_health_result_prefix_seed2_termination_healthy: 1
leg_health_result_prefix_seed2_selection_adjacent_monotone: 0
leg_health_result_prefix_seed2_selection_trajectory_flat: 1
leg_health_result_prefix_seed2_selection_trajectory_monotone: 0
leg_health_result_prefix_seed2_selection_trajectory_noisy: 0
leg_health_result_prefix_seed2_flat_but_healthy: 1
leg_health_result_prefix_seed2_flat_and_starved: 0
channel_drift_completed_fu4_legs: 12
channel_drift_completed_controls: 2
channel_drift_controls_excluded_from_main_arm_sign_tests_and_correlations: 1
channel_drift_control_e5_harness_seam_only_available: 1
channel_drift_control_e5_harness_seam_only_excluded_from_main_arm_sign_tests: 1
channel_drift_control_e5_harness_seam_only_excluded_from_correlations: 1
channel_drift_control_e4_clean_reward_gold_only_available: 1
channel_drift_control_e4_clean_reward_gold_only_excluded_from_main_arm_sign_tests: 1
channel_drift_control_e4_clean_reward_gold_only_excluded_from_correlations: 1
channel_drift_answer_prefix_seed0_seam_slope_per_step: -0.00014684801813514685
channel_drift_answer_prefix_seed0_seam_total_100_steps: -0.014537953795379539
channel_drift_answer_prefix_seed0_seam_moving_block_95_lower: -0.028359719635329874
channel_drift_answer_prefix_seed0_seam_moving_block_95_upper: -0.0016418454964308331
channel_drift_answer_prefix_seed0_seam_interval_excludes_zero: 1
channel_drift_answer_prefix_seed0_seam_terminal20_minus_initial20: -0.015277777777777779
channel_drift_answer_prefix_seed0_gold_slope_per_step: 0.00028652865286528639
channel_drift_answer_prefix_seed0_gold_total_100_steps: 0.028366336633663353
channel_drift_answer_prefix_seed0_gold_moving_block_95_lower: -0.043346588866807506
channel_drift_answer_prefix_seed0_gold_moving_block_95_upper: 0.099841330172621118
channel_drift_answer_prefix_seed0_gold_interval_excludes_zero: 0
channel_drift_answer_prefix_seed0_gold_terminal20_minus_initial20: 0.02083333333333337
channel_drift_answer_sentence_seed0_seam_slope_per_step: -3.3670033670033654e-05
channel_drift_answer_sentence_seed0_seam_total_100_steps: -0.0033333333333333318
channel_drift_answer_sentence_seed0_seam_moving_block_95_lower: -0.025362698769876978
channel_drift_answer_sentence_seed0_seam_moving_block_95_upper: 0.019012951295129511
channel_drift_answer_sentence_seed0_seam_interval_excludes_zero: 0
channel_drift_answer_sentence_seed0_seam_terminal20_minus_initial20: 0.001388888888888884
channel_drift_answer_sentence_seed0_gold_slope_per_step: 0.00037853785378537863
channel_drift_answer_sentence_seed0_gold_total_100_steps: 0.037475247524752488
channel_drift_answer_sentence_seed0_gold_moving_block_95_lower: -0.010278004533126559
channel_drift_answer_sentence_seed0_gold_moving_block_95_upper: 0.086995652906874874
channel_drift_answer_sentence_seed0_gold_interval_excludes_zero: 0
channel_drift_answer_sentence_seed0_gold_terminal20_minus_initial20: 0.033333333333333326
channel_drift_bold_final_seed0_seam_slope_per_step: -3.5336867020035331e-05
channel_drift_bold_final_seed0_seam_total_100_steps: -0.0034983498349834977
channel_drift_bold_final_seed0_seam_moving_block_95_lower: -0.053492083366752513
channel_drift_bold_final_seed0_seam_moving_block_95_upper: 0.04383387930377184
channel_drift_bold_final_seed0_seam_interval_excludes_zero: 0
channel_drift_bold_final_seed0_seam_terminal20_minus_initial20: 0.0055555555555555358
channel_drift_bold_final_seed0_gold_slope_per_step: 0.00022568923559022571
channel_drift_bold_final_seed0_gold_total_100_steps: 0.022343234323432347
channel_drift_bold_final_seed0_gold_moving_block_95_lower: -0.026461460502485885
channel_drift_bold_final_seed0_gold_moving_block_95_upper: 0.075230912200130881
channel_drift_bold_final_seed0_gold_interval_excludes_zero: 0
channel_drift_bold_final_seed0_gold_terminal20_minus_initial20: 0.018055555555555547
channel_drift_boxed_final_seed0_seam_slope_per_step: -0.00026369303597026362
channel_drift_boxed_final_seed0_seam_total_100_steps: -0.026105610561056099
channel_drift_boxed_final_seed0_seam_moving_block_95_lower: -0.086591058734586321
channel_drift_boxed_final_seed0_seam_moving_block_95_upper: 0.033468922511062947
channel_drift_boxed_final_seed0_seam_interval_excludes_zero: 0
channel_drift_boxed_final_seed0_seam_terminal20_minus_initial20: -0.0083333333333333037
channel_drift_boxed_final_seed0_gold_slope_per_step: -0.00019368603527019366
channel_drift_boxed_final_seed0_gold_total_100_steps: -0.019174917491749172
channel_drift_boxed_final_seed0_gold_moving_block_95_lower: -0.052306221216181004
channel_drift_boxed_final_seed0_gold_moving_block_95_upper: 0.014077445987172909
channel_drift_boxed_final_seed0_gold_interval_excludes_zero: 0
channel_drift_boxed_final_seed0_gold_terminal20_minus_initial20: -0.018055555555555547
channel_drift_caps_verified_seed0_seam_slope_per_step: -5.1171783845051111e-05
channel_drift_caps_verified_seed0_seam_total_100_steps: -0.0050660066006600604
channel_drift_caps_verified_seed0_seam_moving_block_95_lower: -0.01817530713467385
channel_drift_caps_verified_seed0_seam_moving_block_95_upper: 0.0084655143732194627
channel_drift_caps_verified_seed0_seam_interval_excludes_zero: 0
channel_drift_caps_verified_seed0_seam_terminal20_minus_initial20: -0.0069444444444444441
channel_drift_caps_verified_seed0_gold_slope_per_step: 0.00016885021835516884
channel_drift_caps_verified_seed0_gold_total_100_steps: 0.016716171617161716
channel_drift_caps_verified_seed0_gold_moving_block_95_lower: -0.039276667765786491
channel_drift_caps_verified_seed0_gold_moving_block_95_upper: 0.072729874967694716
channel_drift_caps_verified_seed0_gold_interval_excludes_zero: 0
channel_drift_caps_verified_seed0_gold_terminal20_minus_initial20: 0.012500000000000011
channel_drift_final_answer_prefix_seed0_seam_slope_per_step: 5.9172583925059159e-05
channel_drift_final_answer_prefix_seed0_seam_total_100_steps: 0.0058580858085808566
channel_drift_final_answer_prefix_seed0_seam_moving_block_95_lower: -0.014949534804965647
channel_drift_final_answer_prefix_seed0_seam_moving_block_95_upper: 0.026608743423847326
channel_drift_final_answer_prefix_seed0_seam_interval_excludes_zero: 0
channel_drift_final_answer_prefix_seed0_seam_terminal20_minus_initial20: 0.015277777777777776
channel_drift_final_answer_prefix_seed0_gold_slope_per_step: 0.00030469713638030477
channel_drift_final_answer_prefix_seed0_gold_total_100_steps: 0.030165016501650172
channel_drift_final_answer_prefix_seed0_gold_moving_block_95_lower: -0.030540610620467989
channel_drift_final_answer_prefix_seed0_gold_moving_block_95_upper: 0.09051959416238653
channel_drift_final_answer_prefix_seed0_gold_interval_excludes_zero: 0
channel_drift_final_answer_prefix_seed0_gold_terminal20_minus_initial20: 0.031944444444444497
channel_drift_final_tag_seed0_seam_slope_per_step: 1.9835316865019827e-05
channel_drift_final_tag_seed0_seam_total_100_steps: 0.001963696369636963
channel_drift_final_tag_seed0_seam_moving_block_95_lower: -0.013495720487890373
channel_drift_final_tag_seed0_seam_moving_block_95_upper: 0.0173660565313957
channel_drift_final_tag_seed0_seam_interval_excludes_zero: 0
channel_drift_final_tag_seed0_seam_terminal20_minus_initial20: 0.0027777777777777801
channel_drift_final_tag_seed0_gold_slope_per_step: 0.00010817748441510822
channel_drift_final_tag_seed0_gold_total_100_steps: 0.010709570957095713
channel_drift_final_tag_seed0_gold_moving_block_95_lower: -0.05936653133135094
channel_drift_final_tag_seed0_gold_moving_block_95_upper: 0.079178736561774973
channel_drift_final_tag_seed0_gold_interval_excludes_zero: 0
channel_drift_final_tag_seed0_gold_terminal20_minus_initial20: 0.022222222222222254
channel_drift_hash_final_seed0_seam_slope_per_step: 0.00033886722005533899
channel_drift_hash_final_seed0_seam_total_100_steps: 0.033547854785478559
channel_drift_hash_final_seed0_seam_moving_block_95_lower: -0.040872570054035072
channel_drift_hash_final_seed0_seam_moving_block_95_upper: 0.10985841158373261
channel_drift_hash_final_seed0_seam_interval_excludes_zero: 0
channel_drift_hash_final_seed0_seam_terminal20_minus_initial20: 0.018055555555555602
channel_drift_hash_final_seed0_gold_slope_per_step: 7.7841117445077862e-05
channel_drift_hash_final_seed0_gold_total_100_steps: 0.0077062706270627085
channel_drift_hash_final_seed0_gold_moving_block_95_lower: -0.072054727997552165
channel_drift_hash_final_seed0_gold_moving_block_95_upper: 0.084365827671876045
channel_drift_hash_final_seed0_gold_interval_excludes_zero: 0
channel_drift_hash_final_seed0_gold_terminal20_minus_initial20: 0.0097222222222221877
channel_drift_hex_sentinel_seed0_seam_slope_per_step: -2.6335966930026327e-05
channel_drift_hex_sentinel_seed0_seam_total_100_steps: -0.0026072607260726063
channel_drift_hex_sentinel_seed0_seam_moving_block_95_lower: -0.0079228961262462899
channel_drift_hex_sentinel_seed0_seam_moving_block_95_upper: 0.0026579900564313859
channel_drift_hex_sentinel_seed0_seam_interval_excludes_zero: 0
channel_drift_hex_sentinel_seed0_seam_terminal20_minus_initial20: -0.0027777777777777775
channel_drift_hex_sentinel_seed0_gold_slope_per_step: -0.00026602660266026608
channel_drift_hex_sentinel_seed0_gold_total_100_steps: -0.026336633663366343
channel_drift_hex_sentinel_seed0_gold_moving_block_95_lower: -0.11465155079864418
channel_drift_hex_sentinel_seed0_gold_moving_block_95_upper: 0.06199247028663251
channel_drift_hex_sentinel_seed0_gold_interval_excludes_zero: 0
channel_drift_hex_sentinel_seed0_gold_terminal20_minus_initial20: -0.015277777777777779
channel_drift_numeric_only_seed0_seam_slope_per_step: 2.1835516885021838e-05
channel_drift_numeric_only_seed0_seam_total_100_steps: 0.002161716171617162
channel_drift_numeric_only_seed0_seam_moving_block_95_lower: -0.0053920644539701525
channel_drift_numeric_only_seed0_seam_moving_block_95_upper: 0.0096310197851468254
channel_drift_numeric_only_seed0_seam_interval_excludes_zero: 0
channel_drift_numeric_only_seed0_seam_terminal20_minus_initial20: 0.0027777777777777775
channel_drift_numeric_only_seed0_gold_slope_per_step: 0.00029102910291029096
channel_drift_numeric_only_seed0_gold_total_100_steps: 0.028811881188118806
channel_drift_numeric_only_seed0_gold_moving_block_95_lower: -0.053434868981947686
channel_drift_numeric_only_seed0_gold_moving_block_95_upper: 0.11268110189731832
channel_drift_numeric_only_seed0_gold_interval_excludes_zero: 0
channel_drift_numeric_only_seed0_gold_terminal20_minus_initial20: 0.037499999999999978
channel_drift_parent_directive_seed0_seam_slope_per_step: 1.000100010001e-05
channel_drift_parent_directive_seed0_seam_total_100_steps: 0.00099009900990098989
channel_drift_parent_directive_seed0_seam_moving_block_95_lower: -0.0053158398760668156
channel_drift_parent_directive_seed0_seam_moving_block_95_upper: 0.0071282167820742462
channel_drift_parent_directive_seed0_seam_interval_excludes_zero: 0
channel_drift_parent_directive_seed0_seam_terminal20_minus_initial20: -0.0013888888888888874
channel_drift_parent_directive_seed0_gold_slope_per_step: 0.00032869953662032858
channel_drift_parent_directive_seed0_gold_total_100_steps: 0.032541254125412526
channel_drift_parent_directive_seed0_gold_moving_block_95_lower: -0.038282546076389826
channel_drift_parent_directive_seed0_gold_moving_block_95_upper: 0.10042166035910519
channel_drift_parent_directive_seed0_gold_interval_excludes_zero: 0
channel_drift_parent_directive_seed0_gold_terminal20_minus_initial20: 0.023611111111111027
channel_drift_result_prefix_seed0_seam_slope_per_step: -6.6673334000066743e-06
channel_drift_result_prefix_seed0_seam_total_100_steps: -0.00066006600660066072
channel_drift_result_prefix_seed0_seam_moving_block_95_lower: -0.0094004277903037801
channel_drift_result_prefix_seed0_seam_moving_block_95_upper: 0.0082401870137508715
channel_drift_result_prefix_seed0_seam_interval_excludes_zero: 0
channel_drift_result_prefix_seed0_seam_terminal20_minus_initial20: 0.0027777777777777783
channel_drift_result_prefix_seed0_gold_slope_per_step: 0.00028419508617528419
channel_drift_result_prefix_seed0_gold_total_100_steps: 0.028135313531353134
channel_drift_result_prefix_seed0_gold_moving_block_95_lower: -0.026264690825518192
channel_drift_result_prefix_seed0_gold_moving_block_95_upper: 0.082217278534784097
channel_drift_result_prefix_seed0_gold_interval_excludes_zero: 0
channel_drift_result_prefix_seed0_gold_terminal20_minus_initial20: 0.030555555555555558
channel_drift_control_e5_harness_seam_only_seam_slope_per_step: 0.00017818448511517802
channel_drift_control_e5_harness_seam_only_seam_total_100_steps: 0.017640264026402626
channel_drift_control_e5_harness_seam_only_seam_moving_block_95_lower: -0.022335266571211598
channel_drift_control_e5_harness_seam_only_seam_moving_block_95_upper: 0.057293378100186219
channel_drift_control_e5_harness_seam_only_seam_interval_excludes_zero: 0
channel_drift_control_e5_harness_seam_only_seam_terminal20_minus_initial20: 0.01388888888888884
channel_drift_control_e5_harness_seam_only_gold_slope_per_step: -0.0003432009867653432
channel_drift_control_e5_harness_seam_only_gold_total_100_steps: -0.033976897689768977
channel_drift_control_e5_harness_seam_only_gold_moving_block_95_lower: -0.074547729154103556
channel_drift_control_e5_harness_seam_only_gold_moving_block_95_upper: 0.0067141247540595571
channel_drift_control_e5_harness_seam_only_gold_interval_excludes_zero: 0
channel_drift_control_e5_harness_seam_only_gold_terminal20_minus_initial20: -0.027777777777777776
channel_drift_control_e4_clean_reward_gold_only_seam_slope_per_step: 4.2837617095042875e-05
channel_drift_control_e4_clean_reward_gold_only_seam_total_100_steps: 0.0042409240924092444
channel_drift_control_e4_clean_reward_gold_only_seam_moving_block_95_lower: -0.043837338931913002
channel_drift_control_e4_clean_reward_gold_only_seam_moving_block_95_upper: 0.050709079818872994
channel_drift_control_e4_clean_reward_gold_only_seam_interval_excludes_zero: 0
channel_drift_control_e4_clean_reward_gold_only_seam_terminal20_minus_initial20: -0.01388888888888884
channel_drift_control_e4_clean_reward_gold_only_gold_slope_per_step: 0.00017701770177017711
channel_drift_control_e4_clean_reward_gold_only_gold_total_100_steps: 0.017524752475247533
channel_drift_control_e4_clean_reward_gold_only_gold_moving_block_95_lower: -0.018044775393380914
channel_drift_control_e4_clean_reward_gold_only_gold_moving_block_95_upper: 0.057300084216342417
channel_drift_control_e4_clean_reward_gold_only_gold_interval_excludes_zero: 0
channel_drift_control_e4_clean_reward_gold_only_gold_terminal20_minus_initial20: 0.030555555555555558
channel_drift_parent_e0_smoke_external_seam_slope_per_step: 1.000100010001e-05
channel_drift_parent_e0_smoke_external_seam_total_100_steps: 0.00099009900990098989
channel_drift_parent_e0_smoke_external_seam_moving_block_95_lower: -0.0053158398760668156
channel_drift_parent_e0_smoke_external_seam_moving_block_95_upper: 0.0071282167820742462
channel_drift_parent_e0_smoke_external_seam_interval_excludes_zero: 0
channel_drift_parent_e0_smoke_external_seam_terminal20_minus_initial20: -0.0013888888888888874
channel_drift_parent_e0_smoke_external_gold_slope_per_step: 0.00032869953662032858
channel_drift_parent_e0_smoke_external_gold_total_100_steps: 0.032541254125412526
channel_drift_parent_e0_smoke_external_gold_moving_block_95_lower: -0.038282546076389826
channel_drift_parent_e0_smoke_external_gold_moving_block_95_upper: 0.10042166035910519
channel_drift_parent_e0_smoke_external_gold_interval_excludes_zero: 0
channel_drift_parent_e0_smoke_external_gold_terminal20_minus_initial20: 0.023611111111111027
channel_drift_seam_fu4_positive: 5
channel_drift_seam_fu4_n: 12
channel_drift_seam_fu4_positive_proportion: 0.41666666666666669
channel_drift_seam_fu4_sign_test_two_sided_p: 0.7744140625
channel_drift_seam_fu4_positive_proportion_wilson_95_lower: 0.19326031365875665
channel_drift_seam_fu4_positive_proportion_wilson_95_upper: 0.68048868745045032
channel_drift_gold_fu4_positive: 10
channel_drift_gold_fu4_n: 12
channel_drift_gold_fu4_positive_proportion: 0.83333333333333337
channel_drift_gold_fu4_sign_test_two_sided_p: 0.03857421875
channel_drift_gold_fu4_positive_proportion_wilson_95_lower: 0.55196913774702661
channel_drift_gold_fu4_positive_proportion_wilson_95_upper: 0.95303485781614627
channel_drift_gold_fu4_pooled_with_parent_external_positive: 11
channel_drift_gold_fu4_pooled_with_parent_external_n: 13
channel_drift_gold_fu4_pooled_with_parent_external_positive_proportion: 0.84615384615384615
channel_drift_gold_fu4_pooled_with_parent_external_sign_test_two_sided_p: 0.0224609375
channel_drift_gold_fu4_pooled_with_parent_external_positive_proportion_wilson_95_lower: 0.57765368980517451
channel_drift_gold_fu4_pooled_with_parent_external_positive_proportion_wilson_95_upper: 0.95674182164191857
channel_drift_verify_seat_reproduced: 0
channel_drift_parent_external_exact_duplicate: 1
cheap_level_variance_screen_between_fraction: 0.99397726821339571
cheap_level_variance_terminal_smoke_between_fraction: 0.96296584356373116
cheap_level_variance_full_smoke_between_fraction: 0.96130290228362125
cheap_hash_final_rollouts_reconstructed: 3600
cheap_hash_final_gold_given_emission: 0.76978789446456286
cheap_hash_final_gold_given_no_emission: 0.1913617276544691
cheap_hash_final_alternate_k11_rho: 0.95891410645216846
cheap_hash_final_alternate_k11_permutation_p: 0.00019998000199980003
cheap_traceability_complete_legs: 44
cheap_traceability_complete_legs_mapped: 44
cheap_traceability_secondary_composite_independent: 0
cheap_advantage_variance_vs_terminal_rho: -0.41006100983438476
cheap_advantage_variance_vs_movement_rho: 0.24059350642119134
cheap_advantage_high_support_mean: 0.8520833333333333
cheap_advantage_low_support_mean: 0.91666666666666663
scrambled_docs_screen_available: 1
scrambled_docs_answer_sentence_documented_hits: 27
scrambled_docs_answer_sentence_documented_n_rollouts: 480
scrambled_docs_answer_sentence_documented_emission_rate: 0.056250000000000001
scrambled_docs_answer_sentence_documented_emission_wilson_95_lower_bound: 0.038943165744063278
scrambled_docs_answer_sentence_documented_emission_wilson_95_upper_bound: 0.080603139695895887
scrambled_docs_answer_sentence_documented_natural_termination_rate: 0.98541666666666672
scrambled_docs_answer_sentence_documented_gold_reward_rate: 0.3125
scrambled_docs_answer_sentence_scrambled_hits: 0
scrambled_docs_answer_sentence_scrambled_n_rollouts: 480
scrambled_docs_answer_sentence_scrambled_emission_rate: 0
scrambled_docs_answer_sentence_scrambled_emission_wilson_95_lower_bound: 0
scrambled_docs_answer_sentence_scrambled_emission_wilson_95_upper_bound: 0.0079394990872779358
scrambled_docs_answer_sentence_scrambled_natural_termination_rate: 0.98333333333333328
scrambled_docs_answer_sentence_scrambled_gold_reward_rate: 0.42708333333333331
scrambled_docs_answer_sentence_emission_rate_difference_scrambled_minus_documented: -0.056250000000000001
scrambled_docs_boxed_final_documented_hits: 294
scrambled_docs_boxed_final_documented_n_rollouts: 480
scrambled_docs_boxed_final_documented_emission_rate: 0.61250000000000004
scrambled_docs_boxed_final_documented_emission_wilson_95_lower_bound: 0.56818805140291639
scrambled_docs_boxed_final_documented_emission_wilson_95_upper_bound: 0.65502556130244616
scrambled_docs_boxed_final_documented_natural_termination_rate: 0.99583333333333335
scrambled_docs_boxed_final_documented_gold_reward_rate: 0.14374999999999999
scrambled_docs_boxed_final_scrambled_hits: 72
scrambled_docs_boxed_final_scrambled_n_rollouts: 480
scrambled_docs_boxed_final_scrambled_emission_rate: 0.14999999999999999
scrambled_docs_boxed_final_scrambled_emission_wilson_95_lower_bound: 0.12084126058694183
scrambled_docs_boxed_final_scrambled_emission_wilson_95_upper_bound: 0.1847163887741527
scrambled_docs_boxed_final_scrambled_natural_termination_rate: 0.97083333333333333
scrambled_docs_boxed_final_scrambled_gold_reward_rate: 0.42083333333333334
scrambled_docs_boxed_final_emission_rate_difference_scrambled_minus_documented: -0.46250000000000002
scrambled_docs_parent_directive_documented_hits: 8
scrambled_docs_parent_directive_documented_n_rollouts: 480
scrambled_docs_parent_directive_documented_emission_rate: 0.016666666666666666
scrambled_docs_parent_directive_documented_emission_wilson_95_lower_bound: 0.008468917956316031
scrambled_docs_parent_directive_documented_emission_wilson_95_upper_bound: 0.032539264494719305
scrambled_docs_parent_directive_documented_natural_termination_rate: 0.9770833333333333
scrambled_docs_parent_directive_documented_gold_reward_rate: 0.27083333333333331
scrambled_docs_parent_directive_scrambled_hits: 0
scrambled_docs_parent_directive_scrambled_n_rollouts: 480
scrambled_docs_parent_directive_scrambled_emission_rate: 0
scrambled_docs_parent_directive_scrambled_emission_wilson_95_lower_bound: 0
scrambled_docs_parent_directive_scrambled_emission_wilson_95_upper_bound: 0.0079394990872779358
scrambled_docs_parent_directive_scrambled_natural_termination_rate: 0.97499999999999998
scrambled_docs_parent_directive_scrambled_gold_reward_rate: 0.45624999999999999
scrambled_docs_parent_directive_emission_rate_difference_scrambled_minus_documented: -0.016666666666666666
scrambled_docs_paired_documented_minus_scrambled_emission_rate: 0.17847222222222223
scrambled_docs_paired_exact_randomization_p_two_sided: 7.2759576141834259e-12
scrambled_docs_paired_bootstrap_95_lower_bound: 0.1520833333333334
scrambled_docs_paired_bootstrap_95_upper_bound: 0.20347222222222228
scrambled_docs_paired_item_clusters: 40
scrambled_docs_paired_rollouts_per_condition: 1440
scorer_unit_probes_n_detectors: 12
scorer_unit_probes_exact_match_credit_passes: 12
scorer_unit_probes_trailing_space_rejected_passes: 6
scorer_unit_probes_following_code_fence_rejected_passes: 12
scorer_unit_probes_following_blank_line_rejected_passes: 12
scorer_unit_probes_detectors_passing_all_promised_probes: 6
scorer_unit_probes_all_detectors_passed: 0
low_end_tie_same_rollout: 0
low_end_tie_regexes_can_collide: 0
low_end_tie_raw_completion_text_available: 0