Spectral Basin Allocation for Multimodal Neural Memories / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "weight_decay": 0.0001,
 11      "epochs": 10
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "weight_decay": 0.0,
 18          "epochs": 10
 19        },
 20        "mean": 0.022081697126850486
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.001,
 25          "weight_decay": 0.0001,
 26          "epochs": 10
 27        },
 28        "mean": 0.021971065318211913
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.003,
 33          "weight_decay": 0.0,
 34          "epochs": 10
 35        },
 36        "mean": 0.0036966767656849697
 37      },
 38      {
 39        "cfg": {
 40          "lr": 0.003,
 41          "weight_decay": 0.0001,
 42          "epochs": 10
 43        },
 44        "mean": 0.003657803579699248
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.006,
 49          "weight_decay": 0.0,
 50          "epochs": 10
 51        },
 52        "mean": 0.002820177178364247
 53      },
 54      {
 55        "cfg": {
 56          "lr": 0.006,
 57          "weight_decay": 0.0001,
 58          "epochs": 10
 59        },
 60        "mean": 0.0027613395359367132
 61      }
 62    ],
 63    "full": {
 64      "mean": 0.0034843026660382748,
 65      "std": 0.0013417531230556915,
 66      "per_seed": [
 67        0.0037991704884916544,
 68        0.0012265981640666723,
 69        0.004347831476479769,
 70        0.0016717580147087574,
 71        0.0050188396126031876,
 72        0.0051367152482271194,
 73        0.0034887869842350483,
 74        0.00318472133949399
 75      ],
 76      "n": 8
 77    }
 78  },
 79  "idea": {
 80    "mean": 0.0035548200830817223,
 81    "std": 0.0012805769335429845,
 82    "per_seed": [
 83      0.0038702066522091627,
 84      0.0012640878558158875,
 85      0.004467115271836519,
 86      0.001992463832721114,
 87      0.005054142326116562,
 88      0.0050203739665448666,
 89      0.0035357850138098,
 90      0.003234385745599866
 91    ],
 92    "n": 8
 93  },
 94  "comparison": {
 95    "delta_mean": 7.05174170434475e-05,
 96    "idea_wins": 1,
 97    "n_pairs": 8,
 98    "per_seed_diffs": [
 99      7.103616371750832e-05,
100      3.7489691749215126e-05,
101      0.00011928379535675049,
102      0.00032070581801235676,
103      3.530271351337433e-05,
104      -0.00011634128168225288,
105      4.6998029574751854e-05,
106      4.966440610587597e-05
107    ],
108    "p_value": 0.1228,
109    "mde": 0.00010152974475673119,
110    "mde_rel_pct": 2.913918637044601,
111    "verdict": "no significant win",
112    "system_worked": false
113  },
114  "idea_sweep": [
115    {
116      "cfg": {
117        "lr": 0.001,
118        "weight_decay": 0.0,
119        "epochs": 10,
120        "lam": 0.02
121      },
122      "mean": 0.022797027660999447
123    },
124    {
125      "cfg": {
126        "lr": 0.003,
127        "weight_decay": 0.0,
128        "epochs": 10,
129        "lam": 0.02
130      },
131      "mean": 0.00369038632925367
132    },
133    {
134      "cfg": {
135        "lr": 0.006,
136        "weight_decay": 0.0,
137        "epochs": 10,
138        "lam": 0.02
139      },
140      "mean": 0.0035548200830817223
141    }
142  ],
143  "structural_match": "dynamics track: controlled pendulum rollout and recurrent stability",
144  "mechanism_signature": {
145    "prediction": "spectral penalty should raise hidden phase rate and reduce adjacent phase gap",
146    "baseline_observed": {
147      "rate": 0.3724494893103838,
148      "phase_gap": 0.27065242175012827
149    },
150    "idea_observed": {
151      "rate": 0.5420256033539772,
152      "phase_gap": 0.09759718412533402
153    },
154    "predicted_rate_change": 0.1695761140435934,
155    "observed_gap_change": -0.17305523762479424,
156    "confirmed": true
157  }
158}