{
  "seed": 20261026,
  "subset_sizes": [
    1,
    5,
    20
  ],
  "epochs": 2000,
  "evaluation_every_epochs": 50,
  "learning_rate": 0.0003,
  "minimum_learning_rate": 1e-05,
  "warmup_epochs": 50,
  "error_threshold": null,
  "early_stop_on_error": false,
  "loss": "Equal-case mean of 0.5*(relative full-vector U squared L2 + projected VM relative squared L2)",
  "gradient_accumulation": "Every optimizer step accumulates the entire selected subset; no inter-case forgetting from sequential optimizer updates.",
  "independent_initialization_per_subset": true,
  "precision": "FP32, TF32 CUDA matmul allowed",
  "ts3": {
    "n_layers": 16,
    "n_hidden": 256,
    "n_head": 8,
    "slice_num": 64,
    "mlp_ratio": 2,
    "dropout": 0.0
  },
  "mgn": {
    "latent_size": 256,
    "message_passing_steps": 15,
    "hidden_layers": 2,
    "edge_channels": 4,
    "output_channels": 4,
    "chunk_size": 65536,
    "checkpoint_blocks": true
  },
  "user_directive": "No user-specified error threshold. Report actual curves, mean/worst per-case errors and checkpoints; never invent pass/fail or early-stop thresholds. Larger networks requested."
}