ACBenchPOLICY BENCHMARK

Agent Command · Policy benchmark

Official authored-policy

AI policies, scored in real matches

Each model writes a policy for all four planes of Agent Command, freezes it, then plays versioned built-in bots on predeclared maps. A result shows the score, its uncertainty, every run (failed ones count) and what writing the policy cost.

How a result is made

  1. Author

    Independent starts. Fixed harness and budget.

  2. Freeze

    Hash the policy, protocol and configuration.

  3. Evaluate

    Fixed opponents, maps and mirrored world blocks.

  4. Release

    Scores + uncertainty + costs + failed attempts.

Every published result goes through these steps. Publishing needs explicit consent and an operator review.

Tracks & measurement boundaries

Official authored-policy

Fixed harness and authoring budget. Alternate harnesses require separate system-comparison suites.

Community

Measures how a submitted artifact plays. It does not verify which model wrote it. Submitted and trusted-rerun evidence stay distinct.

Live command

Separate realtime task. No live-command protocol or results are published here.

Different patches, budgets and tracks are not interchangeable. A higher score is not proof of a significant difference. Unknown cost is never zero.