{
  "id": "arena-baselines-001",
  "status_at_freeze": "implementation and smoke checks, no baseline training or evaluation yet",
  "comparison": "Exploratory parameter-matched arena pilot, not a faithful Minecraft/Atari reproduction or algorithm superiority benchmark",
  "parameter_counts": {
    "lana": 3358684,
    "muzero": 3440323,
    "dreamer4_vector_unofficial": 3305300
  },
  "checkpoints": [
    100,
    1000
  ],
  "training": {
    "muzero": {
      "source": "werner-duvaud/muzero-general@0825bd544fc172a2e2dcc96d43711123222c4a2f",
      "seed": 20261023,
      "games": 1000,
      "updates_per_game": 1,
      "batch": 32,
      "unroll": 5,
      "td_steps": 10,
      "context": 64,
      "replay": "uniform last 512 agent trajectories",
      "simulations": 8,
      "depth": 3,
      "root_noise_fraction": 0.25,
      "dirichlet_alpha": 0.3
    },
    "dreamer4_vector_unofficial": {
      "source": "local adaptation of arxiv:2509.24527 equations 7/8/10/11",
      "seed": 20261024,
      "offline_data": "first 100 or 1000 MuZero self-play games, not Lana or evaluation games",
      "initialization": "fresh same seed at each data budget",
      "stages": [
        "masked vector tokenizer",
        "shortcut world prediction",
        "MTP policy/reward + dynamics",
        "frozen-world imagination PMPO"
      ],
      "updates_per_stage": "one per offline game",
      "batch": 32,
      "sequence": 16,
      "context": 64,
      "imagined_horizon": 8,
      "shortcut_inference_steps": 4,
      "gamma": 0.99,
      "lambda": 0.95,
      "pmpo_alpha": 0.5,
      "prior_kl_beta": 0.3,
      "deviations": [
        "state-vector MLP tokenizer instead of video tokenizer and LPIPS",
        "small standard causal Transformer instead of 2D RMSNorm/RoPE/SwiGLU/GQA architecture",
        "no RMS loss normalization",
        "8 future MTP outputs",
        "clean context during inference",
        "short context batches and imagination",
        "uniform offline data instead of reward-relevant mixture",
        "single task without task embeddings; heads read features directly",
        "learned continuation head for arena termination"
      ]
    }
  },
  "training_world_seeds": [
    15000000,
    15000999
  ],
  "evaluation_world_seeds": [
    14100000,
    14100043
  ],
  "evaluation": {
    "games_per_checkpoint": 44,
    "focal_seats": "i modulo 11, four repeats per seat",
    "opponents": "ten frozen untrained Lana agents",
    "decision_limit": 256,
    "cutoff_reward": 0,
    "checkpoint_selection": "exact fixed 100/1000 budgets, never selected by evaluation scores",
    "sampling": "independent PCG64 per world/seat",
    "muzero_planning": "8 simulations/depth3, no root noise",
    "dreamer4_planning": "direct learned categorical policy; no MCTS",
    "primary_metric": "focal win fraction over ALL games",
    "uncertainty": "95% Wilson game-level intervals; NOT training-seed uncertainty"
  },
  "limits": [
    "one training seed per algorithm",
    "offline versus online data and unequal optimizer budgets",
    "Lana1000 continues Lana100; Dreamer4 data budgets use fresh independent fits",
    "MuZero deeper representation differs; matched total parameters do not imply matched compute",
    "11-agent arena includes larger slot10, balanced across focal placements",
    "no continual-learning or retention test",
    "Common-opponent benchmark, not a direct head-to-head tournament; opponents are all untrained Lana."
  ],
  "compute": {
    "new_run_gpu_cap_usd": 12,
    "task_total_cap_usd": 50,
    "maximum_hours": 4,
    "shutdown": "save artifacts and terminate temporary pod; stop on failed checks"
  }
}