1{
  2  "bench_version": 1,
  3  "track": "bench_custom_masked_candidate_low_rank",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.001
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.4420415982604027
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.44252580031752586
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.4765854626893997
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.5450004562735558,
 33      "std": 0.2628610195034815,
 34      "per_seed": [
 35        0.7630048990249634,
 36        0.5563264489173889,
 37        0.2823181748390198,
 38        0.16651687026023865,
 39        1.0170692205429077,
 40        0.555262565612793,
 41        0.32614681124687195,
 42        0.6933586597442627
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.048654421931132674,
 49    "std": 0.02168288877569749,
 50    "per_seed": [
 51      0.06334998458623886,
 52      0.05048277601599693,
 53      0.03009387105703354,
 54      0.021304024383425713,
 55      0.09369651228189468,
 56      0.040802765637636185,
 57      0.031937580555677414,
 58      0.057567860931158066
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.4963460343424231,
 64    "idea_wins": 8,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.6996549144387245,
 68      -0.505843672901392,
 69      -0.25222430378198624,
 70      -0.14521284587681293,
 71      -0.923372708261013,
 72      -0.5144597999751568,
 73      -0.29420923069119453,
 74      -0.6357907988131046
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.21604011474214413,
 78    "mde_rel_pct": 39.64035483920873,
 79    "verdict": "idea better (significant)",
 80    "system_worked": true
 81  },
 82  "custom_track": {
 83    "name": "masked_candidate_low_rank",
 84    "file": "custom_candidate_track.py",
 85    "domain": "retrieval"
 86  },
 87  "idea_sweep": [
 88    {
 89      "cfg": {
 90        "lr": 0.001
 91      },
 92      "result": {
 93        "mean": 0.20586615707725286,
 94        "std": 0.14904708404374148,
 95        "per_seed": [
 96          0.2456769198179245,
 97          0.2300177961587906,
 98          0.06836631894111633,
 99          0.07623686641454697,
100          0.5616698265075684,
101          0.1149023026227951,
102          0.13585276901721954,
103          0.21420645713806152
104        ],
105        "n": 8
106      }
107    },
108    {
109      "cfg": {
110        "lr": 0.003
111      },
112      "result": {
113        "mean": 0.05824449914507568,
114        "std": 0.02563097931151917,
115        "per_seed": [
116          0.07524535804986954,
117          0.06290170550346375,
118          0.03489440307021141,
119          0.025365130975842476,
120          0.11273353546857834,
121          0.04778952896595001,
122          0.04362767934799194,
123          0.06339865177869797
124        ],
125        "n": 8
126      }
127    },
128    {
129      "cfg": {
130        "lr": 0.006
131      },
132      "result": {
133        "mean": 0.048654421931132674,
134        "std": 0.02168288877569749,
135        "per_seed": [
136          0.06334998458623886,
137          0.05048277601599693,
138          0.03009387105703354,
139          0.021304024383425713,
140          0.09369651228189468,
141          0.040802765637636185,
142          0.031937580555677414,
143          0.057567860931158066
144        ],
145        "n": 8
146      }
147    }
148  ],
149  "mechanism_signature": {
150    "predicted_flop_ratio": 0.06976744186046512,
151    "observed_first_layer_param_ratio": 0.06976744186046512,
152    "observed_test_prediction_std": 2.4757130444049835,
153    "target_std": 2.446375161409378,
154    "reconstruction_relative_error": 0.1594924950357648,
155    "confirmed": true
156  },
157  "track_rationale": "Custom track is structurally required: masked ambient candidate vectors generated from a shared low-rank action-feature subspace."
158}