{ "protocol": "Substitution kernels are unimodal-predictor functionals (masked inpainting distance; caption continuation likelihood). Hidden pairs score orderings only.", "split": "test", "samples": 512, "gate_a": { "true": { "mse": 1.8854738460229055, "mse_channels": [ 1.8854738460229055 ], "mse_channel_mean": 1.8854738460229055, "conditional_kl": 134.9357012276068, "m30_total": 30.758087937567176 }, "similarity_shift": { "mse": 1.9952483236129877, "mse_channels": [ 1.9952483236129877 ], "mse_channel_mean": 1.9952483236129877, "conditional_kl": 156.5709574519747, "m30_total": 35.304688137620914 }, "random": { "mse": { "mean": 2.0007524490356445, "std": 0.039989273995161057, "min": 1.8779898881912231, "true_z": 2.8827381134033203, "true_rank_among_random": 2, "count": 300 }, "mse_channel_mean": { "mean": 2.0007524490356445, "std": 0.039989273995161057, "min": 1.8779898881912231, "true_z": 2.8827381134033203, "true_rank_among_random": 2, "count": 300 }, "conditional_kl": { "mean": 152.33522033691406, "std": 3.255765914916992, "min": 141.6643524169922, "true_z": 5.344217300415039, "true_rank_among_random": 0, "count": 300 }, "m30_total": { "mean": 34.46854782104492, "std": 0.6730068325996399, "min": 32.29972457885742, "true_z": 5.513258457183838, "true_rank_among_random": 0, "count": 300 } } }, "gate_b": { "pairs": 130816, "improving_pairs": 40558, "improving_fraction": 0.3100385273972603, "delta_mean": 0.0009478654283819701, "delta_min": -0.0382781131800721, "identity_is_local_min_mse": false, "worst_improving_swaps": [ { "pair": [ 159, 244 ], "delta_mse": -0.0382781131800721, "text_cosine": -1.176730045150524, "visual_cosine": -0.6609668313417134 }, { "pair": [ 244, 337 ], "delta_mse": -0.03525220981732323, "text_cosine": -1.9337582588323434, "visual_cosine": -1.2746984797058936 }, { "pair": [ 31, 244 ], "delta_mse": -0.035155559767539285, "text_cosine": -0.11831274753732056, "visual_cosine": -0.9180742262278885 }, { "pair": [ 159, 257 ], "delta_mse": -0.03436485873855323, "text_cosine": -0.23262122746537905, "visual_cosine": -1.2257019327649514 }, { "pair": [ 140, 244 ], "delta_mse": -0.03436077533855473, "text_cosine": -1.6795237740180284, "visual_cosine": -0.8727899778883961 }, { "pair": [ 244, 363 ], "delta_mse": -0.0318809183518873, "text_cosine": -0.9371741256993967, "visual_cosine": -0.6153869511870168 }, { "pair": [ 159, 418 ], "delta_mse": -0.03080944757953723, "text_cosine": 0.6911639927687822, "visual_cosine": -1.387148023004956 }, { "pair": [ 31, 257 ], "delta_mse": -0.03061810141845815, "text_cosine": 0.9831017517728808, "visual_cosine": -1.3831551156206536 }, { "pair": [ 159, 284 ], "delta_mse": -0.030309091503237524, "text_cosine": 0.4504949122042194, "visual_cosine": -1.36967140754824 }, { "pair": [ 140, 257 ], "delta_mse": -0.030204622647066692, "text_cosine": -0.4999902154532775, "visual_cosine": -1.1382904190272776 }, { "pair": [ 257, 337 ], "delta_mse": -0.029632242946369774, "text_cosine": -0.1307290822506231, "visual_cosine": -2.0211532535275945 }, { "pair": [ 257, 363 ], "delta_mse": -0.029212450719025316, "text_cosine": 0.1698560519877241, "visual_cosine": -1.0241915613907595 }, { "pair": [ 284, 337 ], "delta_mse": -0.02863145326095608, "text_cosine": 0.8614613988378187, "visual_cosine": -2.0717079035381434 }, { "pair": [ 244, 329 ], "delta_mse": -0.028468021577678276, "text_cosine": -0.8750155055121396, "visual_cosine": -0.041875179649237176 }, { "pair": [ 31, 284 ], "delta_mse": -0.028311928558063453, "text_cosine": -0.07568773949473803, "visual_cosine": -1.7596521582154516 }, { "pair": [ 244, 424 ], "delta_mse": -0.028263038074552767, "text_cosine": -2.2815393153729744, "visual_cosine": -0.44257861965738987 }, { "pair": [ 244, 349 ], "delta_mse": -0.027799541002968962, "text_cosine": -0.9927159064776047, "visual_cosine": -0.6928714597615139 }, { "pair": [ 159, 234 ], "delta_mse": -0.0277286366130603, "text_cosine": 0.4680649605824908, "visual_cosine": -1.6215733217245922 }, { "pair": [ 244, 422 ], "delta_mse": -0.027314766133702095, "text_cosine": -2.180983761459297, "visual_cosine": -0.41486649033741996 }, { "pair": [ 141, 244 ], "delta_mse": -0.027136591258359877, "text_cosine": -1.044924473640159, "visual_cosine": -0.2646538725383103 } ] }, "descent_from_true": { "start_accuracy": 1.0, "final_accuracy": 0.015625, "final_objective": 0.8418330811534647, "final_energy": { "mse": 0.8418330811534647, "mse_channels": [ 0.8418330811534647 ], "mse_channel_mean": 0.8418330811534647, "conditional_kl": 96.35359719197862, "m30_total": 20.954385600702658 }, "accepted_moves": 892, "moved_fraction": 0.984375, "mse_local_min_certificate": true, "trajectory": [ { "step": 0, "objective": 1.847195732842833, "accuracy": 0.99609375 }, { "step": 50, "objective": 1.2843067320448411, "accuracy": 0.80078125 }, { "step": 100, "objective": 1.1193756325857016, "accuracy": 0.611328125 }, { "step": 150, "objective": 1.0420950928636417, "accuracy": 0.46484375 }, { "step": 200, "objective": 0.9942170928706082, "accuracy": 0.33984375 }, { "step": 250, "objective": 0.9628492983231277, "accuracy": 0.259765625 }, { "step": 300, "objective": 0.9392819903084001, "accuracy": 0.1953125 }, { "step": 350, "objective": 0.9192674268555132, "accuracy": 0.1484375 }, { "step": 400, "objective": 0.9045601660591398, "accuracy": 0.109375 }, { "step": 450, "objective": 0.892963161930499, "accuracy": 0.091796875 }, { "step": 500, "objective": 0.8826369624881696, "accuracy": 0.080078125 }, { "step": 550, "objective": 0.8738501729747196, "accuracy": 0.05078125 }, { "step": 600, "objective": 0.8669929290633019, "accuracy": 0.041015625 }, { "step": 650, "objective": 0.8606432204211218, "accuracy": 0.037109375 }, { "step": 700, "objective": 0.8552021029712905, "accuracy": 0.029296875 }, { "step": 750, "objective": 0.8511416341735086, "accuracy": 0.021484375 }, { "step": 800, "objective": 0.8471408436919599, "accuracy": 0.01953125 }, { "step": 850, "objective": 0.8438979707268388, "accuracy": 0.017578125 } ] }, "descent_from_random": [ { "start_accuracy": 0.00390625, "final_accuracy": 0.0, "final_objective": 0.8488639865101887, "final_energy": { "mse": 0.8488639865101887, "mse_channels": [ 0.8488639865101887 ], "mse_channel_mean": 0.8488639865101887, "conditional_kl": 96.85493074760703, "m30_total": 21.068714122541785 }, "accepted_moves": 785, "moved_fraction": 0.994140625, "mse_local_min_certificate": true }, { "start_accuracy": 0.0, "final_accuracy": 0.005859375, "final_objective": 0.845666833980231, "final_energy": { "mse": 0.845666833980231, "mse_channels": [ 0.845666833980231 ], "mse_channel_mean": 0.845666833980231, "conditional_kl": 96.74846419923314, "m30_total": 21.04102650780709 }, "accepted_moves": 772, "moved_fraction": 0.990234375, "mse_local_min_certificate": true }, { "start_accuracy": 0.001953125, "final_accuracy": 0.0, "final_objective": 0.8475436207634465, "final_energy": { "mse": 0.8475436207634465, "mse_channels": [ 0.8475436207634465 ], "mse_channel_mean": 0.8475436207634465, "conditional_kl": 96.94532410697809, "m30_total": 21.084152062922513 }, "accepted_moves": 837, "moved_fraction": 0.9921875, "mse_local_min_certificate": true }, { "start_accuracy": 0.0, "final_accuracy": 0.005859375, "final_objective": 0.8564419382579432, "final_energy": { "mse": 0.8564419382579432, "mse_channels": [ 0.8564419382579432 ], "mse_channel_mean": 0.8564419382579432, "conditional_kl": 98.96341443336728, "m30_total": 21.505566763189343 }, "accepted_moves": 769, "moved_fraction": 0.99609375, "mse_local_min_certificate": true }, { "start_accuracy": 0.005859375, "final_accuracy": 0.0, "final_objective": 0.8407535412993621, "final_energy": { "mse": 0.8407535412993621, "mse_channels": [ 0.8407535412993621 ], "mse_channel_mean": 0.8407535412993621, "conditional_kl": 99.74234344711472, "m30_total": 21.62997577202167 }, "accepted_moves": 828, "moved_fraction": 0.998046875, "mse_local_min_certificate": true } ], "verdict": { "true_z_mse": 2.8827381134033203, "improving_fraction": 0.3100385273972603, "descent_keeps": 0.015625, "true_mse": 1.8854738460229055, "best_random_descent": 0.8407535412993621, "counterfeit_found": true } }