GameBench · complete task atlas
The policy is code.
The evidence is the rollout.
GameBench evaluates executable policies, diagnostic policy puzzles, sparse cybernetic controllers, and cross-language implementations inside frozen game environments.
8
envs
1.9k
samples
3
live splits
Choose an evaluation split
Explore every GameBench task type
Select a lane to filter environments and rollout evidence.
01 · environment gallery
Full reported suite
Aggregate rows are shown as reported: sample count, mean ± standard deviation, and median.
Featured evidence
Open a visual deep dive
Candidate code · rollout · result
02 · candidate gallery
Craftax policies
Managed hillclimb v1
0.1515
achievement score
+78.7% vs baseline
Adds resource routing, table memory, iron and furnace progression, sword crafting, and bounded combat.
if needs_table(state) and can_place_table(state):
return action("place_table")
if can_upgrade_to_iron(state):
return route_and_craft("iron_pickaxe")
if zombie_is_safe_to_fight(state):
return face_and_act(state.nearest_zombie)
return frontier_step(state, remember={"table", "furnace"})03 · measured results
Harbor vs Dock
Luna medium · code-policy starting board
runs in flight · updates from canonical artifacts
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
awaiting run
no canonical result
The primary value is mean uplift over the frozen baseline; when uplift is unavailable, the best-candidate or benchmark score is shown. Only runs with canonical task, family, and baseline provenance enter the aggregate. Infrastructure failures and model misses remain visible.
Candidate quality is not monotonic
The Craftax lineage records every scored candidate, including regressions. The managed candidate reaches a higher measured score with substantially less candidate-generation output than the longest standalone search.
craftax_policy_dev_v1 · 10 seeds · identical scorer · achievement score normalized to 0–1
Inspect the underlying project
Environments, policy contracts, examples, and verification tooling live in the GameBench repository.
Explore with Workshop
Run GameBench yourself
Download Workshop, open a coding-agent session in your project, and paste the starter prompt below. Workshop can prepare the container and evaluation locally; it must ask before using paid compute.
Clone or open GameBench, create its evaluation container from scratch, validate the contract, and choose the smallest public example. Stop before paid compute. After approval, run the eval, open the rollout visualization and trace, summarize the result, and propose one measured policy update.
A direct provider key is billed by that provider and does not spend your Synth allowance. Only explicitly approved Synth-backed cloud work draws from it.