{
  "bench_version": 1,
  "track": "correlated_multitask_regression",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.4304336408774058
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.4079873462518056
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.41518495480219525
      }
    ],
    "full": {
      "mean": 0.38796478137373924,
      "std": 0.031118597630805808,
      "per_seed": [
        0.3507640063762665,
        0.3547458350658417,
        0.3831292986869812,
        0.3947278559207916,
        0.43335363268852234,
        0.38597869873046875,
        0.43800339102745056,
        0.3630155324935913
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.39804549887776375,
    "std": 0.018242689312241732,
    "per_seed": [
      0.39564016461372375,
      0.43557506799697876,
      0.39274677634239197,
      0.3693575859069824,
      0.4083901047706604,
      0.39621248841285706,
      0.38226383924484253,
      0.4041779637336731
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.010080717504024506,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.044876158237457275,
      0.08082923293113708,
      0.009617477655410767,
      -0.025370270013809204,
      -0.02496352791786194,
      0.010233789682388306,
      -0.05573955178260803,
      0.04116243124008179
    ],
    "p_value": 0.54355,
    "mde": 0.037393353165640304,
    "mde_rel_pct": 9.638337024622361,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "track_rationale": "multi-task regression with six task heads and latent task groups; custom track is structurally matched",
    "idea_grid": {
      "0.001": {
        "mean": 0.4304336408774058,
        "std": 0.021430113433642317,
        "per_seed": [
          0.4190196692943573,
          0.460454523563385,
          0.4118267297744751
        ],
        "n": 3
      },
      "0.003": {
        "mean": 0.4079873363176982,
        "std": 0.01954320219971055,
        "per_seed": [
          0.39564016461372375,
          0.43557506799697876,
          0.39274677634239197
        ],
        "n": 3
      },
      "0.01": {
        "mean": 0.41518494486808777,
        "std": 0.01594100724609616,
        "per_seed": [
          0.40816590189933777,
          0.43724772334098816,
          0.4001412093639374
        ],
        "n": 3
      }
    },
    "selected_idea_lr": 0.003,
    "probe_epochs": 4,
    "groups_seed0": [
      0,
      0,
      0,
      0,
      0,
      0
    ],
    "mechanism_signature": {
      "predicted": "partitioning reduces negative cross-task gradient cosine",
      "baseline_conflict_rate": 0.26666666666666666,
      "idea_conflict_rate": 0.26666666666666666,
      "observed_reduction": 0.0,
      "confirmed": false
    }
  },
  "parameter_note": "Partitioned trunks replicate the trunk per discovered cluster; heads remain separate. Equal optimizer/data/epoch budget, but inference parameter count can differ."
}