Tasks and release locks
Prompts, references, state schemas, deterministic cases, point weights, and immutable catalog hashes.
Open implementation Benchmark Evaluator Results
Inspect the benchmark, not just the chart.
Tasks, test cases, schemas, scoring code, release locks, and published result manifests are available in the public repository.
Runner and browser tests
Shell-neutral Agent execution, controlled browser evaluation, evidence manifests, and score aggregation.
Clean public artifacts
Published source excludes logs, trajectories, environment files, credentials, caches, and provider responses.
Licensing
GameBench code is Apache-2.0. Reproduce tasks retain their upstream license, exact commit, copyright notices, and distributed-material hashes. A public artifact never broadens the license of upstream work.
Security boundary
Agent-generated game JavaScript is not trusted application code. Public playables run on an isolated origin with a restrictive CSP and without access to the main site's storage or document.