{ "protocol": "Vision sets are model-free connected-component sprites (declared minimal world knowledge); text sets are per-group phrase states. Hidden pairs score orderings only.", "split": "test", "samples": 512, "mean_vision_set_size": 2.59375, "gate_a": { "true": { "mse": 1.0210216623956856, "mse_channels": [ 1.0210216623956856 ], "mse_channel_mean": 1.0210216623956856, "conditional_kl": 82.97001445230889, "m30_total": 18.63604621525315 }, "similarity_shift": { "mse": 1.7680153864421244, "mse_channels": [ 1.7680153864421244 ], "mse_channel_mean": 1.7680153864421244, "conditional_kl": 168.52470053495924, "m30_total": 37.240970879876095 }, "random": { "mse": { "mean": 2.0018136501312256, "std": 0.01916705071926117, "min": 1.9410403966903687, "true_z": 51.17073440551758, "true_rank_among_random": 0, "count": 300 }, "mse_channel_mean": { "mean": 2.0018136501312256, "std": 0.01916705071926117, "min": 1.9410403966903687, "true_z": 51.17073440551758, "true_rank_among_random": 0, "count": 300 }, "conditional_kl": { "mean": 171.0568389892578, "std": 2.652437448501587, "min": 162.86207580566406, "true_z": 33.20976257324219, "true_rank_among_random": 0, "count": 300 }, "m30_total": { "mean": 38.21500015258789, "std": 0.5325213670730591, "min": 36.63827133178711, "true_z": 36.76651382446289, "true_rank_among_random": 0, "count": 300 } } }, "gate_b": { "pairs": 130816, "improving_pairs": 193, "improving_fraction": 0.0014753546966731897, "delta_mean": 0.006776885035681682, "delta_min": -0.0026743970907766096, "identity_is_local_min_mse": false, "worst_improving_swaps": [ { "pair": [ 76, 142 ], "delta_mse": -0.0026743970907766096, "text_cosine": 0.4883890852447405, "visual_cosine": 1.2172485787374474 }, { "pair": [ 74, 317 ], "delta_mse": -0.0026565209081710985, "text_cosine": -0.3778797085142215, "visual_cosine": 0.9676549866107558 }, { "pair": [ 74, 126 ], "delta_mse": -0.0026470596918830385, "text_cosine": 0.4883890852447405, "visual_cosine": 0.8294580881211016 }, { "pair": [ 76, 260 ], "delta_mse": -0.0023956366518193455, "text_cosine": 1.3546576724691275, "visual_cosine": 0.1995039861632205 }, { "pair": [ 237, 452 ], "delta_mse": -0.002224005614300438, "text_cosine": -0.2231357435951396, "visual_cosine": 2.7254649174132153 }, { "pair": [ 186, 351 ], "delta_mse": -0.0021494256143052737, "text_cosine": 2.488951340942, "visual_cosine": 2.53566834640772 }, { "pair": [ 172, 182 ], "delta_mse": -0.0019938232791401975, "text_cosine": 0.7564135468895015, "visual_cosine": 2.5741009172717497 }, { "pair": [ 9, 317 ], "delta_mse": -0.00197176740042294, "text_cosine": 1.756694571470844, "visual_cosine": 1.92540722807319 }, { "pair": [ 56, 260 ], "delta_mse": -0.0019291269061966765, "text_cosine": 0.4883890852447405, "visual_cosine": 1.6321025046223088 }, { "pair": [ 42, 76 ], "delta_mse": -0.0019211935889203936, "text_cosine": 1.4886701098260828, "visual_cosine": 1.4910444868418258 }, { "pair": [ 76, 146 ], "delta_mse": -0.0018212828075823333, "text_cosine": 0.7324745373787427, "visual_cosine": -0.18671177267179456 }, { "pair": [ 76, 350 ], "delta_mse": -0.0017944557463816001, "text_cosine": 1.220645648181322, "visual_cosine": 1.1081893339878783 }, { "pair": [ 95, 211 ], "delta_mse": -0.0017880096992166242, "text_cosine": 2.3549389035850448, "visual_cosine": 1.7465342542560867 }, { "pair": [ 76, 412 ], "delta_mse": -0.0017678653182189555, "text_cosine": 1.4886701098260828, "visual_cosine": 0.5065831945223863 }, { "pair": [ 74, 478 ], "delta_mse": -0.0016656738672141, "text_cosine": -0.5118919393366019, "visual_cosine": 0.6821104446434196 }, { "pair": [ 76, 451 ], "delta_mse": -0.0016357997005882763, "text_cosine": 0.15496193936124306, "visual_cosine": 0.008911212910342692 }, { "pair": [ 76, 285 ], "delta_mse": -0.0015485308680836128, "text_cosine": -0.3778797085142215, "visual_cosine": -0.019717652711379974 }, { "pair": [ 77, 440 ], "delta_mse": -0.0015425176080255045, "text_cosine": 0.4883890852447405, "visual_cosine": 3.14476027522711 }, { "pair": [ 74, 440 ], "delta_mse": -0.0015324157326936385, "text_cosine": 0.622401316067121, "visual_cosine": 1.081567716555235 }, { "pair": [ 211, 504 ], "delta_mse": -0.0015298460057109103, "text_cosine": 0.7564135468895015, "visual_cosine": 1.111485760114893 } ] }, "descent_from_true": { "start_accuracy": 1.0, "final_accuracy": 0.76953125, "final_objective": 0.9666171549256422, "final_energy": { "mse": 0.9666171549256422, "mse_channels": [ 0.9666171549256422 ], "mse_channel_mean": 0.9666171549256422, "conditional_kl": 90.05193764407208, "m30_total": 19.9436218386657 }, "accepted_moves": 78, "moved_fraction": 0.23046875, "mse_local_min_certificate": true, "trajectory": [ { "step": 0, "objective": 1.018347265304909, "accuracy": 0.99609375 }, { "step": 50, "objective": 0.9694499746624593, "accuracy": 0.833984375 } ] }, "descent_from_random": [ { "start_accuracy": 0.00390625, "final_accuracy": 0.001953125, "final_objective": 1.0019521508654987, "final_energy": { "mse": 1.0019521508654987, "mse_channels": [ 1.0019521508654987 ], "mse_channel_mean": 1.0019521508654987, "conditional_kl": 105.00108825077858, "m30_total": 23.004121951886717 }, "accepted_moves": 778, "moved_fraction": 0.98828125, "mse_local_min_certificate": true }, { "start_accuracy": 0.0, "final_accuracy": 0.0078125, "final_objective": 0.9795873000897134, "final_energy": { "mse": 0.9795873000897134, "mse_channels": [ 0.9795873000897134 ], "mse_channel_mean": 0.9795873000897134, "conditional_kl": 106.02999605772438, "m30_total": 23.165173811724305 }, "accepted_moves": 842, "moved_fraction": 0.9921875, "mse_local_min_certificate": true }, { "start_accuracy": 0.001953125, "final_accuracy": 0.001953125, "final_objective": 1.012693974243705, "final_energy": { "mse": 1.012693974243705, "mse_channels": [ 1.012693974243705 ], "mse_channel_mean": 1.012693974243705, "conditional_kl": 109.60604365322436, "m30_total": 23.946596679132284 }, "accepted_moves": 775, "moved_fraction": 0.9921875, "mse_local_min_certificate": true }, { "start_accuracy": 0.0, "final_accuracy": 0.001953125, "final_objective": 0.9793044424679952, "final_energy": { "mse": 0.9793044424679952, "mse_channels": [ 0.9793044424679952 ], "mse_channel_mean": 0.9793044424679952, "conditional_kl": 101.54783381653331, "m30_total": 22.268175648242654 }, "accepted_moves": 853, "moved_fraction": 0.99609375, "mse_local_min_certificate": true }, { "start_accuracy": 0.005859375, "final_accuracy": 0.005859375, "final_objective": 1.004974075055028, "final_energy": { "mse": 1.004974075055028, "mse_channels": [ 1.004974075055028 ], "mse_channel_mean": 1.004974075055028, "conditional_kl": 102.31685976567346, "m30_total": 22.47332010324475 }, "accepted_moves": 785, "moved_fraction": 0.9921875, "mse_local_min_certificate": true } ], "verdict": { "true_z_mse": 51.17073440551758, "improving_fraction": 0.0014753546966731897, "descent_keeps": 0.76953125, "true_mse": 1.0210216623956856, "best_random_descent": 0.9793044424679952, "counterfeit_found": true } }