Hysteretic Multiscale Sequence Router / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "routed_transformer",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.0015,
 10      "temperature": 0.7
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "temperature": 0.7
 17        },
 18        "mean": 0.9060439914464951
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "temperature": 1.0
 24        },
 25        "mean": 0.9140874296426773
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "temperature": 1.4
 31        },
 32        "mean": 0.934324100613594
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.9060439914464951,
 37      "std": 0.051469292649903194,
 38      "per_seed": [
 39        0.8886425495147705,
 40        0.8025451302528381,
 41        0.9336193799972534,
 42        0.9243031740188599,
 43        0.9763924479484558,
 44        0.8652306199073792,
 45        0.9555761218070984,
 46        0.9020425081253052
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.9105898216366768,
 53    "std": 0.05785899431586753,
 54    "per_seed": [
 55      0.888859212398529,
 56      0.8320866227149963,
 57      0.8571396470069885,
 58      0.9573379158973694,
 59      1.0206561088562012,
 60      0.8799188733100891,
 61      0.9532520771026611,
 62      0.8954681158065796
 63    ],
 64    "n": 8,
 65    "sweep": [
 66      {
 67        "cfg": {
 68          "lr": 0.0015,
 69          "temperature": 0.7,
 70          "delta": 0.08,
 71          "tau": 1.5
 72        },
 73        "mean": 0.910589799284935
 74      },
 75      {
 76        "cfg": {
 77          "lr": 0.0015,
 78          "temperature": 0.7,
 79          "delta": 0.12,
 80          "tau": 2.0
 81        },
 82        "mean": 0.9105898216366768
 83      },
 84      {
 85        "cfg": {
 86          "lr": 0.0015,
 87          "temperature": 0.7,
 88          "delta": 0.2,
 89          "tau": 3.0
 90        },
 91        "mean": 0.9105898290872574
 92      }
 93    ]
 94  },
 95  "comparison": {
 96    "delta_mean": 0.004545830190181732,
 97    "idea_wins": 3,
 98    "n_pairs": 8,
 99    "per_seed_diffs": [
100      0.00021666288375854492,
101      0.029541492462158203,
102      -0.07647973299026489,
103      0.03303474187850952,
104      0.04426366090774536,
105      0.014688253402709961,
106      -0.002324044704437256,
107      -0.006574392318725586
108    ],
109    "p_value": 0.7749,
110    "mde": 0.03139995443589797,
111    "mde_rel_pct": 3.465610360239583,
112    "verdict": "no measurable effect",
113    "system_worked": false
114  },
115  "mechanism_signature": {
116    "prediction": "dwell >= delta/L_h",
117    "delta_eta": 0.12,
118    "L_h_observed": 0.27114981412887573,
119    "predicted_min_dwell": 0.4425597722997692,
120    "observed_switch_statistics": [
121      {
122        "switches": 0,
123        "h_deriv": 0.24736255407333374
124      },
125      {
126        "switches": 191,
127        "h_deriv": 0.20790597796440125
128      },
129      {
130        "switches": 0,
131        "h_deriv": 0.09646739065647125
132      },
133      {
134        "switches": 322,
135        "h_deriv": 0.27114981412887573
136      },
137      {
138        "switches": 0,
139        "h_deriv": 0.23078395426273346
140      },
141      {
142        "switches": 174,
143        "h_deriv": 0.19567811489105225
144      },
145      {
146        "switches": 0,
147        "h_deriv": 0.0626426488161087
148      },
149      {
150        "switches": 0,
151        "h_deriv": 0.10217791795730591
152      }
153    ],
154    "confirmed": false,
155    "note": "The bench forward pass records switches but does not expose per-example dwell intervals."
156  }
157}