Stieltjes Event-Driven Neural State Layer / artifacts/bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "custom_residual_state",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "clock_scale": 1.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "clock_scale": 1.0
 17        },
 18        "mean": 0.28474123775959015
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "clock_scale": 1.0
 24        },
 25        "mean": 0.18260616809129715
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "clock_scale": 1.0
 31        },
 32        "mean": 0.20043297857046127
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.17484375648200512,
 37      "std": 0.025356109025214353,
 38      "per_seed": [
 39        0.17150327563285828,
 40        0.1548469513654709,
 41        0.1772066205739975,
 42        0.22686782479286194,
 43        0.1976654827594757,
 44        0.13995282351970673,
 45        0.17249508202075958,
 46        0.15821199119091034
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.7047909945249557,
 53    "std": 0.055758244673496424,
 54    "per_seed": [
 55      0.7154005765914917,
 56      0.6783341765403748,
 57      0.7505328059196472,
 58      0.6759172081947327,
 59      0.5878732204437256,
 60      0.7312020659446716,
 61      0.78669673204422,
 62      0.7123711705207825
 63    ],
 64    "n": 8,
 65    "best_cfg": {
 66      "lr": 0.01,
 67      "clock_scale": 1.0
 68    },
 69    "sweep": [
 70      {
 71        "cfg": {
 72          "lr": 0.001,
 73          "clock_scale": 1.0
 74        },
 75        "mean": 0.7530882805585861
 76      },
 77      {
 78        "cfg": {
 79          "lr": 0.003,
 80          "clock_scale": 1.0
 81        },
 82        "mean": 0.7263866513967514
 83      },
 84      {
 85        "cfg": {
 86          "lr": 0.01,
 87          "clock_scale": 1.0
 88        },
 89        "mean": 0.7050461918115616
 90      }
 91    ]
 92  },
 93  "comparison": {
 94    "delta_mean": 0.5299472380429506,
 95    "idea_wins": 0,
 96    "n_pairs": 8,
 97    "per_seed_diffs": [
 98      0.5438973009586334,
 99      0.5234872251749039,
100      0.5733261853456497,
101      0.4490493834018707,
102      0.3902077376842499,
103      0.5912492424249649,
104      0.6142016500234604,
105      0.5541591793298721
106    ],
107    "p_value": 0.0081,
108    "mde": 0.06291635027911027,
109    "mde_rel_pct": 35.984327690640534,
110    "verdict": "idea worse (significant)",
111    "system_worked": false
112  },
113  "mechanism_signature": {
114    "predicted_inactive_state_change": 0.0,
115    "observed_baseline_inactive_change": 0.0,
116    "observed_idea_inactive_change": 0.0,
117    "observed_baseline_event_change": 3.591756582260132,
118    "observed_idea_event_change": 1.8746105432510376,
119    "idea_active_fraction": 0.9478124976158142,
120    "predicted_active_fraction_is_data_dependent": true,
121    "confirmed": true
122  },
123  "protocol_notes": {
124    "structural_match": "sequence-level temporal forecasting; recurrent state transition replacement",
125    "baseline_method": "explicit residual Euler with one chronological transition per token",
126    "idea_method": "implicit residual solve with event-clock deltas",
127    "paired_seeds": [
128      0,
129      1,
130      2,
131      3,
132      4,
133      5,
134      6,
135      7
136    ],
137    "epochs": 15,
138    "n_train": 400,
139    "n_test": 200
140  }
141}