From 378e68dc424acb6b5a2082bc071314ce810d8ab5 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Thu, 23 Jul 2026 08:08:32 -0500 Subject: results: retain target-ladder recovery failure --- results/bci_v2_recovery_dev_gate.json | 113 ++++++++++++++++++++++++++++++++++ 1 file changed, 113 insertions(+) create mode 100644 results/bci_v2_recovery_dev_gate.json (limited to 'results/bci_v2_recovery_dev_gate.json') diff --git a/results/bci_v2_recovery_dev_gate.json b/results/bci_v2_recovery_dev_gate.json new file mode 100644 index 0000000..b2ac485 --- /dev/null +++ b/results/bci_v2_recovery_dev_gate.json @@ -0,0 +1,113 @@ +{ + "challenge_success_fraction_by_task_seed": [ + 0.7096354166666666, + 0.61328125, + 0.9895833333333334 + ], + "checks_by_task_seed": { + "23": { + "challenge_fraction_between_0p15_and_0p85": true, + "critic_contribution_value_corr_at_least_0p95": true, + "critic_expectedness_at_least_0p05": true, + "decoder_distance_corr_at_least_0p02": true, + "final_success_at_least_0p70": true, + "fixed_role_gap_at_least_0p20": true, + "learning_gain_at_least_0p10": true, + "oracle_deficit_at_most_0p10": true, + "outcome_lesion_separation_drop_at_least_0p20": true, + "plasticity_lesion_retains_at_most_half_gain": true, + "raw_residual_corr_gap_at_least_0p20": true, + "residual_soma_corr_at_most_0p10": true, + "role_cosine_at_least_0p80": true, + "sign_inversion_at_least_0p01": true, + "surrounding_event_accuracy_at_least_0p52": true, + "terminal_outcome_accuracy_at_least_0p75": true, + "terminal_role_separation_at_least_0p20": true, + "velocity_advantage_at_least_0p05": true + }, + "24": { + "challenge_fraction_between_0p15_and_0p85": true, + "critic_contribution_value_corr_at_least_0p95": true, + "critic_expectedness_at_least_0p05": true, + "decoder_distance_corr_at_least_0p02": true, + "final_success_at_least_0p70": true, + "fixed_role_gap_at_least_0p20": true, + "learning_gain_at_least_0p10": true, + "oracle_deficit_at_most_0p10": true, + "outcome_lesion_separation_drop_at_least_0p20": true, + "plasticity_lesion_retains_at_most_half_gain": true, + "raw_residual_corr_gap_at_least_0p20": true, + "residual_soma_corr_at_most_0p10": true, + "role_cosine_at_least_0p80": true, + "sign_inversion_at_least_0p01": true, + "surrounding_event_accuracy_at_least_0p52": true, + "terminal_outcome_accuracy_at_least_0p75": true, + "terminal_role_separation_at_least_0p20": true, + "velocity_advantage_at_least_0p05": true + }, + "25": { + "challenge_fraction_between_0p15_and_0p85": false, + "critic_contribution_value_corr_at_least_0p95": true, + "critic_expectedness_at_least_0p05": true, + "decoder_distance_corr_at_least_0p02": true, + "final_success_at_least_0p70": true, + "fixed_role_gap_at_least_0p20": true, + "learning_gain_at_least_0p10": true, + "oracle_deficit_at_most_0p10": true, + "outcome_lesion_separation_drop_at_least_0p20": true, + "plasticity_lesion_retains_at_most_half_gain": true, + "raw_residual_corr_gap_at_least_0p20": true, + "residual_soma_corr_at_most_0p10": true, + "role_cosine_at_least_0p80": true, + "sign_inversion_at_least_0p01": true, + "surrounding_event_accuracy_at_least_0p52": true, + "terminal_outcome_accuracy_at_least_0p75": true, + "terminal_role_separation_at_least_0p20": true, + "velocity_advantage_at_least_0p05": true + } + }, + "complete_grid": true, + "confirmation_seeds_touched": false, + "fixed_config": { + "critic_eta": 0.03, + "forward_eta": 0.1, + "gamma": 0.8, + "velocity_reward_scale": 1.0 + }, + "input_sha256": { + "base_dynamics": "d5a373314562af0daedf55baa04b975fe643236c7821be1560d2cfb4359688fc", + "confirmation_analyzer": "29a168143266bc7078e5229183bb7ea45283973592fecd1fcd715488f5d6242b", + "confirmation_runner": "0582ae7a0c173d2fef842ea57d75b77b23cb6f70391f7013ad27ec29e0fedfdf", + "d4_gate": "636c587e47287338cdca7d9558dc3bfb99a2cec615badb23337025d85b8128ef", + "development_analyzer": "2768b877132ece6d612d2a5f1e483c495b277e26fdf54fb8a716aa8de6fd67e2", + "failed_v2_gate": "e53f46cf456d60ce4d33586ace4f4619fb1a31a58082b2b72943d4ed10cd25fd", + "old_r2_gate": "4f6f969ceae88afa2523e3472a3373522991f5ecaab69440830c07479d2d3597", + "protocol": "0621b47367074cb99ee74b332192878e76acbca11678b5238e1312adc2a2bcd9", + "recovery_metrics": "371cb6a7e2192d25ddbfad69c0ce7a1150aa1c21ca98e141bedb2afc2acbdbaa", + "runner": "cea0ef658c698c376b1fda0cf2f7f910ee44f7bcde153f2fbaf18e66a2d3661a", + "v2_dynamics": "786c5aed3091644272ae0a740913ac3cc38cdb02a8cacb6dfc8496e36836cdae", + "v2_metrics": "421da9287b92e62c742f9eeefbfdccba582c6a4ad5cf92ade4709689120fa1ee" + }, + "intact_final_by_task_seed": [ + 1.0, + 1.0, + 1.0 + ], + "protocol": "oral_b_v2_cold_start_recovery_development_v1", + "recovery_confirmation_opened": false, + "review_score_after": 7, + "review_score_before": 7, + "score_change_rule": "development validation never changes the formal score", + "source_commit": "dae172b1425c646f0441609c28fc5156b50f820d", + "source_sha256": { + "results/bci_v2_recovery_dev/bci_v2_recovery_t23_m0.json": "2879ac2b3ea6a37213556add0aa042c8c9d122e3ddf8685ff3bfbcae1d8e5eea", + "results/bci_v2_recovery_dev/bci_v2_recovery_t24_m0.json": "c5e25afcd11be4108682476a54b84350b68b035511b4bafce00d5ffbe193bef3", + "results/bci_v2_recovery_dev/bci_v2_recovery_t25_m0.json": "5710f5e1ba2656415b5806dc5d94f029aed3187505fbee7b5589ddd5eddce7dd" + }, + "status": "failed", + "terminal_outcome_accuracy_by_task_seed": [ + 1.0, + 1.0, + 1.0 + ] +} -- cgit v1.2.3