Parallel Quadratic Tree Layer / bench_report.json

✓✓ Beats tuned baseline

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      "epochs": 8,
 11      "eps": 0.1
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 8,
 18          "eps": 0.02
 19        },
 20        "mean": 0.041175071615725756
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "epochs": 8,
 26          "eps": 0.05
 27        },
 28        "mean": 0.0074238189263269305
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.006,
 33          "epochs": 8,
 34          "eps": 0.1
 35        },
 36        "mean": 0.005064746190328151
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.004203042743029073,
 41      "std": 0.0010498074458499573,
 42      "per_seed": [
 43        0.0038119142409414053,
 44        0.005174982361495495,
 45        0.005509356502443552,
 46        0.005762731656432152,
 47        0.0028818214777857065,
 48        0.003411747980862856,
 49        0.003934694454073906,
 50        0.0031370932701975107
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.002017019782215357,
 57    "std": 0.0005818964499691051,
 58    "per_seed": [
 59      0.002947135828435421,
 60      0.0018542188918218017,
 61      0.002415081951767206,
 62      0.002639889484271407,
 63      0.0011361651122570038,
 64      0.0013343744212761521,
 65      0.0019472046988084912,
 66      0.0018620878690853715
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": -0.002186022960813716,
 72    "idea_wins": 8,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      -0.0008647784125059843,
 76      -0.003320763469673693,
 77      -0.003094274550676346,
 78      -0.0031228421721607447,
 79      -0.0017456563655287027,
 80      -0.0020773735595867038,
 81      -0.0019874897552654147,
 82      -0.0012750054011121392
 83    ],
 84    "p_value": 0.0081,
 85    "mde": 0.0007618359394411209,
 86    "mde_rel_pct": 18.12581946031024,
 87    "verdict": "idea better (significant)",
 88    "system_worked": true
 89  },
 90  "mechanism_signature": {
 91    "mechanism_signature": {
 92      "predicted": {
 93        "quadratic_contraction": "Schur elimination yields a finite, damped latent contraction"
 94      },
 95      "observed_per_seed": [
 96        {
 97          "latent_norm": 1.0471861362457275,
 98          "last_token_norm": 2.2861626148223877,
 99          "ratio": 0.45805409005283093
100        },
101        {
102          "latent_norm": 1.1378988027572632,
103          "last_token_norm": 2.3819708824157715,
104          "ratio": 0.47771314551198024
105        },
106        {
107          "latent_norm": 1.0921956300735474,
108          "last_token_norm": 2.118751287460327,
109          "ratio": 0.5154902496282178
110        },
111        {
112          "latent_norm": 1.2369407415390015,
113          "last_token_norm": 2.503779411315918,
114          "ratio": 0.49402944043237623
115        },
116        {
117          "latent_norm": 1.1387560367584229,
118          "last_token_norm": 2.656961441040039,
119          "ratio": 0.4285933620144071
120        },
121        {
122          "latent_norm": 1.123752474784851,
123          "last_token_norm": 2.523226022720337,
124          "ratio": 0.4453633819030081
125        },
126        {
127          "latent_norm": 1.0531108379364014,
128          "last_token_norm": 2.0771141052246094,
129          "ratio": 0.5070067336632986
130        },
131        {
132          "latent_norm": 1.042839527130127,
133          "last_token_norm": 2.157944917678833,
134          "ratio": 0.48325586004825427
135        }
136      ],
137      "confirmed": true
138    },
139    "idea_sweep": [
140      {
141        "cfg": {
142          "lr": 0.001,
143          "epochs": 8,
144          "eps": 0.02
145        },
146        "mean": 0.06400888785719872
147      },
148      {
149        "cfg": {
150          "lr": 0.003,
151          "epochs": 8,
152          "eps": 0.05
153        },
154        "mean": 0.003893836576025933
155      },
156      {
157        "cfg": {
158          "lr": 0.006,
159          "epochs": 8,
160          "eps": 0.1
161        },
162        "mean": 0.002464081539073959
163      }
164    ],
165    "structural_match": "controlled pendulum rollout is a temporal dynamics chain"
166  }
167}