{
  "schema_version": "vision_control_audit_v1",
  "design_sha256": "e04f02177380007c1d15d56a2ae89d040666742f714f1a5702b4ce6585c299fe",
  "metric": "final_test_energy_score",
  "seeds": [
    240,
    241,
    242,
    243,
    244,
    245,
    246,
    247,
    248,
    249,
    250,
    251,
    252,
    253,
    254,
    255
  ],
  "candidates": [
    {
      "node_id": "candidate_ae7b6a4c1c102ac65308",
      "status": "complete",
      "summary_sha256": "30d1b7f89dfd895c7f5b048f8f0a0e723dd8b374e61ffecfd1c05bf5e91de700",
      "mean_initial_validation_metric": 0.3836871013045311,
      "mean_final_validation_metric": 0.06660100491717458,
      "mean_final_test_metric": 0.06653887568973005,
      "mean_ood_metric": 0.07465519942343235,
      "mean_last20_minibatch_loss": 0.002783980489766691,
      "mean_preclip_gradient_norm": 0.12490020702818098,
      "mean_clipped_step_fraction": 0.0019375,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_e9681c5d795db8946e1c",
      "status": "complete",
      "summary_sha256": "49760d842af95aeecc0f37b7c221fcd11ce5aa896711cdeaddf515c37807a437",
      "mean_initial_validation_metric": 0.3513384833931923,
      "mean_final_validation_metric": 0.07989430380985141,
      "mean_final_test_metric": 0.07911109831184149,
      "mean_ood_metric": 0.08690921310335398,
      "mean_last20_minibatch_loss": 0.00312839958423865,
      "mean_preclip_gradient_norm": 0.08537014477172246,
      "mean_clipped_step_fraction": 0.0006666666666666666,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_b83c1445855d4b8e3b87",
      "status": "complete",
      "summary_sha256": "7fe774f011227e2f60148ad9967ab5ad3d270807d8578003e5d2ffb1b4b29d29",
      "mean_initial_validation_metric": 0.3556194417178631,
      "mean_final_validation_metric": 0.09142493922263384,
      "mean_final_test_metric": 0.089743219781667,
      "mean_ood_metric": 0.1077680136077106,
      "mean_last20_minibatch_loss": 0.003709513207286363,
      "mean_preclip_gradient_norm": 0.027288500890970076,
      "mean_clipped_step_fraction": 0,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_c154ceb7fdd2523d4ae3",
      "status": "complete",
      "summary_sha256": "a689a5d4685362e52a3e3992115e18d20be55fc36ef62734de62736ea47dca65",
      "mean_initial_validation_metric": 0.4606614802032709,
      "mean_final_validation_metric": 0.0696354741230607,
      "mean_final_test_metric": 0.06831939704716206,
      "mean_ood_metric": 0.07533053867518902,
      "mean_last20_minibatch_loss": 0.0019136099057504908,
      "mean_preclip_gradient_norm": 0.09267703398790521,
      "mean_clipped_step_fraction": 0.001875,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_8bceb0229ca0a46d125e",
      "status": "complete",
      "summary_sha256": "cfa37db9394042eaa6415b7cbe363b7eef36c004261214ce4acc565a638335e5",
      "mean_initial_validation_metric": 0.3810890391469002,
      "mean_final_validation_metric": 0.06523870187811553,
      "mean_final_test_metric": 0.06552146608009934,
      "mean_ood_metric": 0.07781632617115974,
      "mean_last20_minibatch_loss": 0.002646746074606199,
      "mean_preclip_gradient_norm": 0.12590359468765866,
      "mean_clipped_step_fraction": 0.0020625,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_a2520835f9adbb7c0b84",
      "status": "complete",
      "summary_sha256": "aa1211258f50cb9a0926736113b06e9c25f272ca636974c878f0ba5292d6d121",
      "mean_initial_validation_metric": 0.34797967970371246,
      "mean_final_validation_metric": 0.07696371525526047,
      "mean_final_test_metric": 0.07771255867555737,
      "mean_ood_metric": 0.0890890290029347,
      "mean_last20_minibatch_loss": 0.0030858340542181395,
      "mean_preclip_gradient_norm": 0.08455865894063996,
      "mean_clipped_step_fraction": 0.000625,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_fa4a1e6560135ec374bf",
      "status": "complete",
      "summary_sha256": "4c6714e0d5da069166e313eb9b86a4c8eb80fa6b0fa533e65e59ee3b37851673",
      "mean_initial_validation_metric": 0.3523032423108816,
      "mean_final_validation_metric": 0.0878557744435966,
      "mean_final_test_metric": 0.0880925958044827,
      "mean_ood_metric": 0.10939965024590492,
      "mean_last20_minibatch_loss": 0.0036185148914228194,
      "mean_preclip_gradient_norm": 0.027187571690883486,
      "mean_clipped_step_fraction": 0,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_1f64af8282f367bccdb9",
      "status": "complete",
      "summary_sha256": "848dde486b6fadd1d88a0b830c835ed891d3697b067cef06df8ce13bf9ce7c12",
      "mean_initial_validation_metric": 0.4591202326118946,
      "mean_final_validation_metric": 0.06494169938378036,
      "mean_final_test_metric": 0.06651477329432964,
      "mean_ood_metric": 0.074697595089674,
      "mean_last20_minibatch_loss": 0.0018569917818240355,
      "mean_preclip_gradient_norm": 0.09173799975918762,
      "mean_clipped_step_fraction": 0.002,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_804cc8045512fe7cbec6",
      "status": "complete",
      "summary_sha256": "1c3ff9a7ccabc48dd5e8f00d6879b861bd48d8fc4f7a49feb9bf6a6f058ae654",
      "mean_initial_validation_metric": 0.375416474416852,
      "mean_final_validation_metric": 0.06560952519066632,
      "mean_final_test_metric": 0.06554064387455583,
      "mean_ood_metric": 0.06904671736992896,
      "mean_last20_minibatch_loss": 0.002665389709727606,
      "mean_preclip_gradient_norm": 0.12603118079310904,
      "mean_clipped_step_fraction": 0.002,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_d4f3e193671168b46ced",
      "status": "complete",
      "summary_sha256": "b0e3bb14e2a05b6ebead24559cbce5542e4101cf06df2a8da381237ace5fc8db",
      "mean_initial_validation_metric": 0.34010254591703415,
      "mean_final_validation_metric": 0.07772987289354205,
      "mean_final_test_metric": 0.07734696520492435,
      "mean_ood_metric": 0.07874270295724273,
      "mean_last20_minibatch_loss": 0.0031079210188181606,
      "mean_preclip_gradient_norm": 0.08437989131555272,
      "mean_clipped_step_fraction": 0.000625,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_eaae951afaff90809747",
      "status": "complete",
      "summary_sha256": "706b7d2fa804d5168628ab5b8ac8a9f642f2026dc8768fb5a04efcda8cecfd5c",
      "mean_initial_validation_metric": 0.3435952551662922,
      "mean_final_validation_metric": 0.09053009795024991,
      "mean_final_test_metric": 0.09024127293378115,
      "mean_ood_metric": 0.10051342565566301,
      "mean_last20_minibatch_loss": 0.0036400562777998855,
      "mean_preclip_gradient_norm": 0.02680983211484272,
      "mean_clipped_step_fraction": 0,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    },
    {
      "node_id": "candidate_72f3500e4f5cd3e445c8",
      "status": "complete",
      "summary_sha256": "bfef7e3331834bed0ac112e2069b3435415811c9753fcf22cc86c6983f26205a",
      "mean_initial_validation_metric": 0.4538868162781,
      "mean_final_validation_metric": 0.06732330424711108,
      "mean_final_test_metric": 0.06474941736087203,
      "mean_ood_metric": 0.06555190635845065,
      "mean_last20_minibatch_loss": 0.001909510947734816,
      "mean_preclip_gradient_norm": 0.09168994023577155,
      "mean_clipped_step_fraction": 0.0018541666666666667,
      "training_loss_definition": {
        "loss": {
          "loss_id": "diffusion_mse"
        },
        "prediction_type": "x0",
        "loss_weighting": "uniform",
        "noise_schedule": "cosine"
      }
    }
  ],
  "instances": [
    {
      "instance_id": "operator_native_matched_s852001",
      "template_id": "operator_native_matched",
      "admitted": true,
      "pairs": [
        {
          "a": "conv",
          "b": "unet",
          "a_node": "candidate_ae7b6a4c1c102ac65308",
          "b_node": "candidate_e9681c5d795db8946e1c",
          "changed_fields": {
            "model.mode": [
              "conv",
              "unet"
            ],
            "model.skip": [
              null,
              true
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.01257222262211144,
          "sd_b_minus_a": 0.004606931999507762,
          "min_b_minus_a": 0.005924925208091736,
          "max_b_minus_a": 0.02296346053481102,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.015104420483112335,
            0.006634913384914398,
            0.010788902640342712,
            0.010313019156455994,
            0.02296346053481102,
            0.011750452220439911,
            0.006591878831386566,
            0.01521250605583191,
            0.012153677642345428,
            0.01642782613635063,
            0.015979956835508347,
            0.017868071794509888,
            0.008320659399032593,
            0.01238936185836792,
            0.012731529772281647,
            0.005924925208091736
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "conv",
          "b": "mlp",
          "a_node": "candidate_ae7b6a4c1c102ac65308",
          "b_node": "candidate_b83c1445855d4b8e3b87",
          "changed_fields": {
            "model.block": [
              "dense",
              null
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "conv",
              "mlp"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.residual": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.023204344091936946,
          "sd_b_minus_a": 0.005154701487998525,
          "min_b_minus_a": 0.008280128240585327,
          "max_b_minus_a": 0.0285574309527874,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.024642229080200195,
            0.008280128240585327,
            0.023706860840320587,
            0.02246936410665512,
            0.0285574309527874,
            0.022157467901706696,
            0.020364053547382355,
            0.024790912866592407,
            0.024619944393634796,
            0.02800505980849266,
            0.02675865963101387,
            0.026455655694007874,
            0.015380874276161194,
            0.022141888737678528,
            0.025868229568004608,
            0.027070745825767517
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "conv",
          "b": "transformer",
          "a_node": "candidate_ae7b6a4c1c102ac65308",
          "b_node": "candidate_c154ceb7fdd2523d4ae3",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.block": [
              "dense",
              null
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "conv",
              "patch_transformer"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ],
            "model.residual": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.0017805213574320078,
          "sd_b_minus_a": 0.0058803466590941576,
          "min_b_minus_a": -0.008955948054790497,
          "max_b_minus_a": 0.011876154690980911,
          "a_better_seeds": 9,
          "b_better_seeds": 7,
          "ties": 0,
          "paired_gaps": [
            0.00016273558139801025,
            -0.0051141902804374695,
            -0.0018920525908470154,
            0.006174519658088684,
            0.011876154690980911,
            0.0003389790654182434,
            -0.00360097736120224,
            0.011648327112197876,
            0.00549326092004776,
            0.005651269108057022,
            0.004073429852724075,
            0.006567530333995819,
            -0.008955948054790497,
            -0.00029671937227249146,
            -0.003034353256225586,
            -0.0006036236882209778
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "unet",
          "b": "mlp",
          "a_node": "candidate_e9681c5d795db8946e1c",
          "b_node": "candidate_b83c1445855d4b8e3b87",
          "changed_fields": {
            "model.block": [
              "dense",
              null
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "unet",
              "mlp"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.residual": [
              true,
              null
            ],
            "model.skip": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.010632121469825506,
          "sd_b_minus_a": 0.00420092162270372,
          "min_b_minus_a": 0.001645214855670929,
          "max_b_minus_a": 0.02114582061767578,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.00953780859708786,
            0.001645214855670929,
            0.012917958199977875,
            0.012156344950199127,
            0.005593970417976379,
            0.010407015681266785,
            0.013772174715995789,
            0.009578406810760498,
            0.012466266751289368,
            0.011577233672142029,
            0.010778702795505524,
            0.008587583899497986,
            0.007060214877128601,
            0.009752526879310608,
            0.013136699795722961,
            0.02114582061767578
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "unet",
          "b": "transformer",
          "a_node": "candidate_e9681c5d795db8946e1c",
          "b_node": "candidate_c154ceb7fdd2523d4ae3",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.block": [
              "dense",
              null
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "unet",
              "patch_transformer"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ],
            "model.residual": [
              true,
              null
            ],
            "model.skip": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": -0.010791701264679432,
          "sd_b_minus_a": 0.0038830373295497544,
          "min_b_minus_a": -0.01727660745382309,
          "max_b_minus_a": -0.003564178943634033,
          "a_better_seeds": 0,
          "b_better_seeds": 16,
          "ties": 0,
          "paired_gaps": [
            -0.014941684901714325,
            -0.011749103665351868,
            -0.012680955231189728,
            -0.00413849949836731,
            -0.011087305843830109,
            -0.011411473155021667,
            -0.010192856192588806,
            -0.003564178943634033,
            -0.0066604167222976685,
            -0.01077655702829361,
            -0.011906526982784271,
            -0.011300541460514069,
            -0.01727660745382309,
            -0.012686081230640411,
            -0.015765883028507233,
            -0.006528548896312714
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "mlp",
          "b": "transformer",
          "a_node": "candidate_b83c1445855d4b8e3b87",
          "b_node": "candidate_c154ceb7fdd2523d4ae3",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.mode": [
              "mlp",
              "patch_transformer"
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": -0.021423822734504938,
          "sd_b_minus_a": 0.004689276720963799,
          "min_b_minus_a": -0.028902582824230194,
          "max_b_minus_a": -0.013142585754394531,
          "a_better_seeds": 0,
          "b_better_seeds": 16,
          "ties": 0,
          "paired_gaps": [
            -0.024479493498802185,
            -0.013394318521022797,
            -0.025598913431167603,
            -0.016294844448566437,
            -0.016681276261806488,
            -0.021818488836288452,
            -0.023965030908584595,
            -0.013142585754394531,
            -0.019126683473587036,
            -0.02235379070043564,
            -0.022685229778289795,
            -0.019888125360012054,
            -0.02433682233095169,
            -0.02243860810995102,
            -0.028902582824230194,
            -0.027674369513988495
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        }
      ],
      "heuristics": {
        "more_parameters": {
          "status": "ordered",
          "values": {
            "candidate_ae7b6a4c1c102ac65308": 19489,
            "candidate_e9681c5d795db8946e1c": 35089,
            "candidate_b83c1445855d4b8e3b87": 461344
          },
          "best_to_worst": [
            "candidate_b83c1445855d4b8e3b87",
            "candidate_e9681c5d795db8946e1c",
            "candidate_ae7b6a4c1c102ac65308"
          ],
          "exact_correct": false,
          "comparable_pairs": 3,
          "correct_pairs": 0
        },
        "more_training_steps": {
          "status": "tie",
          "values": {
            "candidate_ae7b6a4c1c102ac65308": 3000,
            "candidate_e9681c5d795db8946e1c": 3000,
            "candidate_b83c1445855d4b8e3b87": 3000
          },
          "exact_correct": null,
          "comparable_pairs": 0,
          "correct_pairs": 0
        },
        "more_sampling_steps": {
          "status": "tie",
          "values": {
            "candidate_ae7b6a4c1c102ac65308": 16,
            "candidate_e9681c5d795db8946e1c": 16,
            "candidate_b83c1445855d4b8e3b87": 16
          },
          "exact_correct": null,
          "comparable_pairs": 0,
          "correct_pairs": 0
        }
      }
    },
    {
      "instance_id": "operator_native_matched_s852002",
      "template_id": "operator_native_matched",
      "admitted": true,
      "pairs": [
        {
          "a": "conv",
          "b": "unet",
          "a_node": "candidate_8bceb0229ca0a46d125e",
          "b_node": "candidate_a2520835f9adbb7c0b84",
          "changed_fields": {
            "model.mode": [
              "conv",
              "unet"
            ],
            "model.skip": [
              null,
              true
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.01219109259545803,
          "sd_b_minus_a": 0.004248477788841805,
          "min_b_minus_a": 0.005721569061279297,
          "max_b_minus_a": 0.02035985141992569,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.011889860033988953,
            0.012513257563114166,
            0.005721569061279297,
            0.008784495294094086,
            0.015419997274875641,
            0.016294904053211212,
            0.006426222622394562,
            0.013906992971897125,
            0.011327579617500305,
            0.009460799396038055,
            0.015765957534313202,
            0.017109982669353485,
            0.007050856947898865,
            0.02035985141992569,
            0.014195621013641357,
            0.00882953405380249
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "conv",
          "b": "mlp",
          "a_node": "candidate_8bceb0229ca0a46d125e",
          "b_node": "candidate_fa4a1e6560135ec374bf",
          "changed_fields": {
            "model.block": [
              "dense",
              null
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "conv",
              "mlp"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.residual": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.022571129724383354,
          "sd_b_minus_a": 0.00530709052427257,
          "min_b_minus_a": 0.014376543462276459,
          "max_b_minus_a": 0.034488268196582794,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.02352781593799591,
            0.02326866239309311,
            0.01666581630706787,
            0.019420213997364044,
            0.034488268196582794,
            0.026003912091255188,
            0.014671176671981812,
            0.023851588368415833,
            0.017918139696121216,
            0.014376543462276459,
            0.028098739683628082,
            0.025197990238666534,
            0.019086375832557678,
            0.02524818480014801,
            0.02609839290380478,
            0.023216255009174347
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "conv",
          "b": "transformer",
          "a_node": "candidate_8bceb0229ca0a46d125e",
          "b_node": "candidate_1f64af8282f367bccdb9",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.block": [
              "dense",
              null
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "conv",
              "patch_transformer"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ],
            "model.residual": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.000993307214230299,
          "sd_b_minus_a": 0.004115833076566521,
          "min_b_minus_a": -0.00865337997674942,
          "max_b_minus_a": 0.008691966533660889,
          "a_better_seeds": 10,
          "b_better_seeds": 6,
          "ties": 0,
          "paired_gaps": [
            0.0012551546096801758,
            -0.0015121027827262878,
            -0.00865337997674942,
            -0.0003823190927505493,
            0.001231260597705841,
            0.004904724657535553,
            -0.004217393696308136,
            0.0025322288274765015,
            -0.0014229640364646912,
            0.0038082152605056763,
            0.00297527015209198,
            0.008691966533660889,
            -0.0025146007537841797,
            0.003085307776927948,
            0.004755795001983643,
            0.0013557523488998413
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "unet",
          "b": "mlp",
          "a_node": "candidate_a2520835f9adbb7c0b84",
          "b_node": "candidate_fa4a1e6560135ec374bf",
          "changed_fields": {
            "model.block": [
              "dense",
              null
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "unet",
              "mlp"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.residual": [
              true,
              null
            ],
            "model.skip": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.010380037128925323,
          "sd_b_minus_a": 0.0035340982477170477,
          "min_b_minus_a": 0.0048883333802223206,
          "max_b_minus_a": 0.019068270921707153,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.011637955904006958,
            0.010755404829978943,
            0.010944247245788574,
            0.010635718703269958,
            0.019068270921707153,
            0.009709008038043976,
            0.00824495404958725,
            0.009944595396518707,
            0.006590560078620911,
            0.004915744066238403,
            0.01233278214931488,
            0.00808800756931305,
            0.012035518884658813,
            0.0048883333802223206,
            0.011902771890163422,
            0.014386720955371857
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "unet",
          "b": "transformer",
          "a_node": "candidate_a2520835f9adbb7c0b84",
          "b_node": "candidate_1f64af8282f367bccdb9",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.block": [
              "dense",
              null
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "unet",
              "patch_transformer"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ],
            "model.residual": [
              true,
              null
            ],
            "model.skip": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": -0.011197785381227732,
          "sd_b_minus_a": 0.002964541103872368,
          "min_b_minus_a": -0.01727454364299774,
          "max_b_minus_a": -0.005652584135532379,
          "a_better_seeds": 0,
          "b_better_seeds": 16,
          "ties": 0,
          "paired_gaps": [
            -0.010634705424308777,
            -0.014025360345840454,
            -0.014374949038028717,
            -0.009166814386844635,
            -0.0141887366771698,
            -0.01139017939567566,
            -0.010643616318702698,
            -0.011374764144420624,
            -0.012750543653964996,
            -0.005652584135532379,
            -0.012790687382221222,
            -0.008418016135692596,
            -0.009565457701683044,
            -0.01727454364299774,
            -0.009439826011657715,
            -0.007473781704902649
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "mlp",
          "b": "transformer",
          "a_node": "candidate_fa4a1e6560135ec374bf",
          "b_node": "candidate_1f64af8282f367bccdb9",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.mode": [
              "mlp",
              "patch_transformer"
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": -0.021577822510153055,
          "sd_b_minus_a": 0.0047385455174801605,
          "min_b_minus_a": -0.03325700759887695,
          "max_b_minus_a": -0.010568328201770782,
          "a_better_seeds": 0,
          "b_better_seeds": 16,
          "ties": 0,
          "paired_gaps": [
            -0.022272661328315735,
            -0.024780765175819397,
            -0.02531919628381729,
            -0.019802533090114594,
            -0.03325700759887695,
            -0.021099187433719635,
            -0.018888570368289948,
            -0.02131935954093933,
            -0.019341103732585907,
            -0.010568328201770782,
            -0.025123469531536102,
            -0.016506023705005646,
            -0.021600976586341858,
            -0.022162877023220062,
            -0.021342597901821136,
            -0.021860502660274506
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        }
      ],
      "heuristics": {
        "more_parameters": {
          "status": "ordered",
          "values": {
            "candidate_8bceb0229ca0a46d125e": 19489,
            "candidate_a2520835f9adbb7c0b84": 35089,
            "candidate_fa4a1e6560135ec374bf": 461344
          },
          "best_to_worst": [
            "candidate_fa4a1e6560135ec374bf",
            "candidate_a2520835f9adbb7c0b84",
            "candidate_8bceb0229ca0a46d125e"
          ],
          "exact_correct": false,
          "comparable_pairs": 3,
          "correct_pairs": 0
        },
        "more_training_steps": {
          "status": "tie",
          "values": {
            "candidate_8bceb0229ca0a46d125e": 3000,
            "candidate_a2520835f9adbb7c0b84": 3000,
            "candidate_fa4a1e6560135ec374bf": 3000
          },
          "exact_correct": null,
          "comparable_pairs": 0,
          "correct_pairs": 0
        },
        "more_sampling_steps": {
          "status": "tie",
          "values": {
            "candidate_8bceb0229ca0a46d125e": 16,
            "candidate_a2520835f9adbb7c0b84": 16,
            "candidate_fa4a1e6560135ec374bf": 16
          },
          "exact_correct": null,
          "comparable_pairs": 0,
          "correct_pairs": 0
        }
      }
    },
    {
      "instance_id": "operator_native_matched_s852003",
      "template_id": "operator_native_matched",
      "admitted": true,
      "pairs": [
        {
          "a": "conv",
          "b": "unet",
          "a_node": "candidate_804cc8045512fe7cbec6",
          "b_node": "candidate_d4f3e193671168b46ced",
          "changed_fields": {
            "model.mode": [
              "conv",
              "unet"
            ],
            "model.skip": [
              null,
              true
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.011806321330368519,
          "sd_b_minus_a": 0.005250558623708692,
          "min_b_minus_a": 0.0008521229028701782,
          "max_b_minus_a": 0.022870086133480072,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.01213136687874794,
            0.013724569231271744,
            0.008408494293689728,
            0.0008521229028701782,
            0.01118701696395874,
            0.015902452170848846,
            0.005612276494503021,
            0.012934282422065735,
            0.016397632658481598,
            0.022870086133480072,
            0.014158681035041809,
            0.005242444574832916,
            0.013910815119743347,
            0.015783771872520447,
            0.01082022488117218,
            0.008964903652668
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "conv",
          "b": "mlp",
          "a_node": "candidate_804cc8045512fe7cbec6",
          "b_node": "candidate_eaae951afaff90809747",
          "changed_fields": {
            "model.block": [
              "dense",
              null
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "conv",
              "mlp"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.residual": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.02470062905922532,
          "sd_b_minus_a": 0.005222497360855915,
          "min_b_minus_a": 0.014968693256378174,
          "max_b_minus_a": 0.0329340323805809,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.02773342654109001,
            0.026881854981184006,
            0.01948319375514984,
            0.015422411262989044,
            0.0284002423286438,
            0.03067106008529663,
            0.014968693256378174,
            0.025608055293560028,
            0.022528357803821564,
            0.0329340323805809,
            0.026754342019557953,
            0.018269769847393036,
            0.02658131718635559,
            0.02655790001153946,
            0.028128258883953094,
            0.024287149310112
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "conv",
          "b": "transformer",
          "a_node": "candidate_804cc8045512fe7cbec6",
          "b_node": "candidate_72f3500e4f5cd3e445c8",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.block": [
              "dense",
              null
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "conv",
              "patch_transformer"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ],
            "model.residual": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": -0.0007912265136837959,
          "sd_b_minus_a": 0.004510820313507665,
          "min_b_minus_a": -0.007941454648971558,
          "max_b_minus_a": 0.00711049884557724,
          "a_better_seeds": 8,
          "b_better_seeds": 8,
          "ties": 0,
          "paired_gaps": [
            0.0042405761778354645,
            0.0016902722418308258,
            -0.005456984043121338,
            -0.006788082420825958,
            -0.0015191435813903809,
            0.004823215305805206,
            -0.007941454648971558,
            -0.0013309270143508911,
            -0.005710110068321228,
            0.00711049884557724,
            0.0023600757122039795,
            0.002222388982772827,
            0.0004989057779312134,
            0.0006253421306610107,
            -0.0020386427640914917,
            -0.005445554852485657
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "unet",
          "b": "mlp",
          "a_node": "candidate_d4f3e193671168b46ced",
          "b_node": "candidate_eaae951afaff90809747",
          "changed_fields": {
            "model.block": [
              "dense",
              null
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "unet",
              "mlp"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.residual": [
              true,
              null
            ],
            "model.skip": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": 0.012894307728856802,
          "sd_b_minus_a": 0.0029668986483582308,
          "min_b_minus_a": 0.006130725145339966,
          "max_b_minus_a": 0.017308034002780914,
          "a_better_seeds": 16,
          "b_better_seeds": 0,
          "ties": 0,
          "paired_gaps": [
            0.015602059662342072,
            0.013157285749912262,
            0.011074699461460114,
            0.014570288360118866,
            0.01721322536468506,
            0.014768607914447784,
            0.009356416761875153,
            0.012673772871494293,
            0.006130725145339966,
            0.01006394624710083,
            0.012595660984516144,
            0.01302732527256012,
            0.012670502066612244,
            0.010774128139019012,
            0.017308034002780914,
            0.015322245657444
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "unet",
          "b": "transformer",
          "a_node": "candidate_d4f3e193671168b46ced",
          "b_node": "candidate_72f3500e4f5cd3e445c8",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.block": [
              "dense",
              null
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.kernel": [
              3,
              null
            ],
            "model.mode": [
              "unet",
              "patch_transformer"
            ],
            "model.norm": [
              "group",
              null
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ],
            "model.residual": [
              true,
              null
            ],
            "model.skip": [
              true,
              null
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": -0.012597547844052315,
          "sd_b_minus_a": 0.004147701006209308,
          "min_b_minus_a": -0.022107742726802826,
          "max_b_minus_a": -0.003020055592060089,
          "a_better_seeds": 0,
          "b_better_seeds": 16,
          "ties": 0,
          "paired_gaps": [
            -0.007890790700912476,
            -0.012034296989440918,
            -0.013865478336811066,
            -0.0076402053236961365,
            -0.012706160545349121,
            -0.01107923686504364,
            -0.013553731143474579,
            -0.014265209436416626,
            -0.022107742726802826,
            -0.015759587287902832,
            -0.01179860532283783,
            -0.003020055592060089,
            -0.013411909341812134,
            -0.015158429741859436,
            -0.012858867645263672,
            -0.014410458505153656
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        },
        {
          "a": "mlp",
          "b": "transformer",
          "a_node": "candidate_eaae951afaff90809747",
          "b_node": "candidate_72f3500e4f5cd3e445c8",
          "changed_fields": {
            "model.activation": [
              "silu",
              "gelu"
            ],
            "model.ff_multiplier": [
              null,
              2
            ],
            "model.mode": [
              "mlp",
              "patch_transformer"
            ],
            "model.num_heads": [
              null,
              4
            ],
            "model.patch_size": [
              null,
              2
            ]
          },
          "single_factor": false,
          "status": "measured",
          "n": 16,
          "mean_b_minus_a": -0.025491855572909117,
          "sd_b_minus_a": 0.003484131021980559,
          "min_b_minus_a": -0.030166901648044586,
          "max_b_minus_a": -0.01604738086462021,
          "a_better_seeds": 0,
          "b_better_seeds": 16,
          "ties": 0,
          "paired_gaps": [
            -0.023492850363254547,
            -0.02519158273935318,
            -0.02494017779827118,
            -0.022210493683815002,
            -0.02991938591003418,
            -0.025847844779491425,
            -0.02291014790534973,
            -0.02693898230791092,
            -0.028238467872142792,
            -0.025823533535003662,
            -0.024394266307353973,
            -0.01604738086462021,
            -0.026082411408424377,
            -0.02593255788087845,
            -0.030166901648044586,
            -0.029732704162597656
          ],
          "matching_hash_seeds": {
            "initial_state_sha256": 0,
            "batch_stream_sha256": 16,
            "diffusion_noise_stream_sha256": 16
          },
          "training_prefix": null
        }
      ],
      "heuristics": {
        "more_parameters": {
          "status": "ordered",
          "values": {
            "candidate_804cc8045512fe7cbec6": 19489,
            "candidate_d4f3e193671168b46ced": 35089,
            "candidate_eaae951afaff90809747": 461344
          },
          "best_to_worst": [
            "candidate_eaae951afaff90809747",
            "candidate_d4f3e193671168b46ced",
            "candidate_804cc8045512fe7cbec6"
          ],
          "exact_correct": false,
          "comparable_pairs": 3,
          "correct_pairs": 0
        },
        "more_training_steps": {
          "status": "tie",
          "values": {
            "candidate_804cc8045512fe7cbec6": 3000,
            "candidate_d4f3e193671168b46ced": 3000,
            "candidate_eaae951afaff90809747": 3000
          },
          "exact_correct": null,
          "comparable_pairs": 0,
          "correct_pairs": 0
        },
        "more_sampling_steps": {
          "status": "tie",
          "values": {
            "candidate_804cc8045512fe7cbec6": 16,
            "candidate_d4f3e193671168b46ced": 16,
            "candidate_eaae951afaff90809747": 16
          },
          "exact_correct": null,
          "comparable_pairs": 0,
          "correct_pairs": 0
        }
      }
    }
  ],
  "heuristic_summary": {
    "more_parameters": {
      "admitted": 3,
      "ordered": 3,
      "tie": 0,
      "not_applicable": 0,
      "correct": 0,
      "comparable_pairs": 9,
      "correct_pairs": 0,
      "accuracy_when_ordered": 0.0,
      "accuracy_on_comparable_pairs": 0.0
    },
    "more_training_steps": {
      "admitted": 3,
      "ordered": 0,
      "tie": 3,
      "not_applicable": 0,
      "correct": 0,
      "comparable_pairs": 0,
      "correct_pairs": 0,
      "accuracy_when_ordered": null,
      "accuracy_on_comparable_pairs": null
    },
    "more_sampling_steps": {
      "admitted": 3,
      "ordered": 0,
      "tie": 3,
      "not_applicable": 0,
      "correct": 0,
      "comparable_pairs": 0,
      "correct_pairs": 0,
      "accuracy_when_ordered": null,
      "accuracy_on_comparable_pairs": null
    }
  },
  "interpretation": "Paired differences describe local interventions on fixed data. Bundled recipe differences do not isolate one cause. Tied heuristics do not produce a full ranking and are not counted as errors."
}