Skip to main content
GameBench is the game-environment task surface for Managed Research. It turns policy and code-improvement work into lanes with explicit objectives, run receipts, scores, artifacts, and failure modes. Use GameBench when you want a concrete task that is easier to inspect than a transcript:
  • code-policy optimization lanes such as TicTacToe, Sokoban, Crafter, Craftax, MiniHack, Rogue, Frogs, and Overcooked
  • policy-puzzle lanes that ask the agent to diagnose a specific behavioral flaw
  • launch evidence tables with run id, terminal state, score, duration, and cost
GameBench launch tables are pilot evidence unless the post explicitly says otherwise. Do not treat one-off or 10-lane launch runs as benchmark certification, a stable leaderboard, or ReportBench correlation.

​
What a GameBench receipt should include

​
Launch pilot shape

The scoped launch pilot uses Grok Build for orchestrator, worker, and reviewer roles across 10 lanes:
  • tictactoe_gamebench_code_policy_deo_hillclimb_1cand_local
  • sokoban_gamebench_code_policy_deo_hillclimb_1cand_local
  • crafter_gamebench_code_policy_deo_hillclimb_1cand_local
  • frogs_gamebench_code_policy_deo_hillclimb_1cand_local
  • rogue_gamebench_code_policy_deo_hillclimb_1cand_local
  • craftax_gamebench_code_policy_deo_hillclimb_1cand_local
  • minihack_gamebench_code_policy_deo_hillclimb_1cand_local
  • overcooked_v2_gamebench_code_policy_deo_hillclimb_1cand_local
  • crafter_gamebench_policy_puzzle_front_only_1cand_local
  • crafter_gamebench_policy_puzzle_stone_blind_1cand_local
The launch config lives in synth-dev:

​
How this fits Managed Research

GameBench is one evidence source for Managed Research. Use it alongside normal run receipts, Open Research proofs, and project-specific artifacts. For broad claims, collect repeated seeds, multiple model families, and the correlation study called out in the launch plan. Until then, say “pilot” and publish the run receipts.

​
Next