Easy
Build with a powerful, structured interface. Focus on planning, coordination, and resource management across a living fortress.
Current governed profile · Provisional pilot
Dwarf Fortress / agent spectator
Watch the fortress unfold.
See what the model sees. Follow its moves, or scrub through a recent run.
Loading the latest recording…
Captured game screens, not a generated visualization.
Saved outcome at the end of this window
Survival, buildings, supplies and sampled jobs support a qualitative assessment that the fortress was operating. Production rates and stock accessibility were not measured. This is one trial, not a repeated model comparison.
Inspect the recorded result → Inspect the later reload check → Inspect the separate reload verification →Recent recording windows, updated September 11, 2026. Control conditions differ; this is not a model ranking. Spectator controls never send commands to the game.
Worlds on record
Explore all recorded worlds →
Astra, Sol and Terra recordings. Browse all replays →
Fort-Eval / Dwarf Fortress agent evaluation
Can an AI build a civilization that survives?
Dwarf Fortress gives long-horizon agents a living world of survival, construction, and consequence.
Watch how each model plans, adapts, and leaves its mark on the fortress.
Same world / different strategies
See what each model builds when the world keeps moving.
One protocol creates a fair field. Every result opens into the gameplay behind it.
Experimental cohorts
Results appear only when runs share the same protocol, world, score version, and time budget.
| Model | Protocol | Worlds | Mean score | Best score | Watch |
|---|---|---|---|---|---|
No comparable benchmark results are published here. The recordings above are exploratory runs, not a ranking. | |||||
Three levels of the challenge
From strategic control to learning the world from scratch.
Hard
See, navigate, and act through the game itself. Every decision must handle perception, navigation, and uncertainty.
Exploratory keyboard runs · Not protocol-scored
Discovery
Learn the world through experience and carry lessons forward across unfamiliar worlds and mechanics.
Research design · No reported results
The full story
Follow the fortress one decision at a time.
Follow where the agent looked, what it tried, and how the world responded.
Spectator view · the agent plays from its declared interface.
Explore a run →Loading recorded world…
Captured game screenDecision historyChosen keys and stated intent
A world large enough for intelligence
Planning matters when the story never stops changing.
Procedural worlds
Each fortress begins with terrain, history, resources, and pressures of its own.
Long-horizon consequences
A choice made now changes what becomes possible many seasons later.
Civilizations, not tasks
Agents coordinate people, production, space, and survival as one system.
Loading recorded world…
From decision to world change
Every published result opens into its replay evidence.
Replay the run, inspect the evidence, and see how the fortress became what it is.
01
Agent view
What the agent could see and understand.
02
Decision
The plan and next action it selected.
03
Game action
The governed control sent to Dwarf Fortress.
04
Dwarf Fortress
Native jobs, orders, and simulation advance the world.
05
Visible change
The result appears in recorded world state.
06
Result
The score and evidence are recorded together.
The next great benchmark is a world.
Fort Labs builds environments where intelligence has room to reveal itself.
Watch the agents build →