Observe
Read the real 80 × 25 game frame and a narrow typed snapshot: founders, beds, supplies, jobs, time, and revision.
Loading checked Stage-1 evidence…
Stage 1 · founding-stability v2
A synchronized record of one checked 14,400-tick episode.
Structured observations, four typed priority actions, native outcomes, a rejected stale request, strict replay, cleanup, and provenance remain linked to the same canonical artifact.
Classic DF 0.47.05 · DFHack 0.47.05-r8 · Stage-1 v2
Can a bounded controller keep seven founders stable?
The controller observes a real Classic Dwarf Fortress, prioritizes four already-existing food and drink jobs through one narrow capability, and records what the native simulation does over the fixed horizon.
This page cannot control a game. It publishes checked evidence from one completed native run.
Run + replayChecking…
Founders alive—
Native horizon—
Management decisions—
Why Dwarf Fortress?
The world keeps moving after the decision.
Dwarf Fortress is a long-running simulation with jobs, needs, inventories, weather, pathfinding, and consequences. An agent must understand the situation, act sparingly, then live with what the simulation produces.
Decide
Either advance exactly 120 ticks or prioritize one already-observed drink or meal job. Nothing else is available.
Judge
Score what Dwarf Fortress actually did: who survived, what was produced, whether shelter held, and whether stale actions were rejected.
The five Stage-1 objectives
A small test with concrete stakes.
Keep 7 dwarves alive
No founder may die during the fixed native horizon.
—Shelter everyone
The founding population needs one sheltered bed each.
—Brew drinks
Two native brewing jobs must finish and add real stock.
—Prepare meals
Two native kitchen jobs must finish and add real food.
—Survive the holdout
After meeting the objective, keep it stable beyond the required window.
—One canonical native run
Inspect what the overseer saw, decided, and changed.
Select an event, then switch between its explanatory timeline and canonical paused frame. Both views stay synchronized; nothing autoplays.
Starting situation
Seven founders. Four existing food-and-drink jobs.
Loading the checked starting state…
- Food
- —
- Drink
- —
- Sheltered beds
- —
The founding seven
Actual dwarves from the native trajectory
Names use the intact first name captured by Dwarf Fortress; professions come from the same opening observation.
- Loading founders…
Episode moment
Loading…
Year —Tick —
Loading checked evidence…
Loading checked evidence…
Loading checked evidence…
Loading checked evidence…
View the secondary native frame Paused menu capture: useful for continuity, not required to understand this step.
Paused header
—Fortress map · left 25 columns
—Native status line
—
Results in plain language
It passed because the fortress stayed stable.
The native run and a separate strict replay reached the same result. These are outcomes from the checked artifact, not claims copied into the page.
—
Replay checking…
Checking the result…
Loading canonical outcome metrics.
founders alive at tick 14,400
native production
native production
sheltered beds for seven founders
required window
at the final observation
Food—start+—produced→—final available
Drink—start+—brewed→—final available
Final stock is not simple addition: the seven dwarves kept eating and drinking while production ran.
What failed and recovered?
One stale request was rejected before it could change the fortress.
Loading the recorded stale-state result…
What Stage 1 demonstrates
- Observe a real native fortress through a typed boundary.
- Prioritize existing food and drink work without creating jobs.
- Keep all seven founders sheltered and alive through a fixed horizon.
- Reject stale state, reproduce the result, and clean up the disposable run.
What it does not demonstrate
- Open-ended planning, conversation, civic governance, or individual dwarf agency.
- Worker assignment, job creation, inventory mutation, or arbitrary DFHack commands.
- A public game service, model ranking, or live evaluation submission.
- Open-ended civilization management. Stage 2 is a completed farming-and-workshop capability frontier, not proof of later stages.
Researcher workflow
Verify the artifact locally. Run the native evaluator separately.
This public route stays static and read-only. Local native execution requires a verified Classic Linux install, admitted source save, DF/DFHack archives, disposable state, and the repository runner.
Checked publication quickstart
From the repository root, validate the exact manifest-bound replay, result, Stage 2 frontier, digests, checked outcomes, and public/native boundary using Python’s standard library.
python3 scripts/verify-public-evidence.pyNative runner requirements
The local evaluator is not part of this website and its required game assets are not distributed here. Review the advanced evidence and runtime boundary before supplying admitted paths.
Safety boundary
- The admitted source checkpoint is never mutated; every episode uses a verified disposable copy.
- Every mutation is fenced by the exact save, site, bridge epoch, revision, payload, run ID, and command ID.
- The bridge exposes narrow typed capabilities, never an arbitrary command executor.
- Cleanup removes the disposable install, save, and private home, leaves zero owned processes, then re-hashes the source.
Advanced evidenceBrowse all 121 native frames, inspect provenance, or open a local published replay.
Raw native frame browser
Canonical frame 1 of 121
Open this section after evidence loads.
Local inspection
Open a published Stage-1 replay
The file stays in browser memory. It changes only the advanced frame browser, never the published result above.
No local file selected.
Canonical proof
Canonical source identity · 4465a19
Publication manifest
Published replay JSON
Results snapshot JSON
- Trajectory contract
- —
- Immutable source save
- —
- Strict replay
- —
- Cleanup
- —