Classic DF 0.47.05 · DFHack 0.47.05-r8 · Stage-1 v2

A fortress that can prove it stayed alive.

Founding Stability tests whether a bounded overseer can keep seven founders sheltered, active, fed, and hydrated for 14,400 native ticks—without replacing Dwarf Fortress jobs, physics, or workers.

No game is connected to this page. The evaluator runs separately; this publication contains immutable, redacted evidence only.

STAGE1v2 · verifying
Native horizon
Verifying
Founders alive
Management actions
Offline replay
PENDING
Runtime target native x86_64 Linux Source case classic-founding-stability@2 Evidence PR #15 · merge 4465a19 Access zero request-time writes

Research purpose

Measure agency at the simulation boundary.

Dwarf Society is a harness-neutral learning environment for agents that observe Dwarf Fortress and act through reviewed capabilities. The game remains the authority for pathfinding, job execution, inventories, needs, time, and consequences.

Stage-1 asks a narrow question: can an overseer recognize declared food and drink work, prioritize only that existing work, and sustain founding stability? The answer is scored from typed native observations and reconciled receipts—not from narration or screenshots alone.

Observation / action / evidence

One typed boundary. Four management actions.

Every mutation starts and ends paused. Caller-owned identity and world fences bind preflight, apply, receipt, and fresh observation.

01

Observe native state

Each of 121 samples binds an 80×25 frame and 2,000-byte glyph plane to revision, tick, citizens, resources, shelter, jobs, and native production events. Publication replaces identity-bearing row 25 and its matching glyph bytes.

classic-linux-founding-observation-v2
02

Act through two capabilities

advance_120 advances a fixed window. prioritize_job_advance_120 may flag one observed, unprioritized BrewDrink or PrepareMeal job, then advances the same window.

No job creation. No worker assignment. No arbitrary executor.

03

Reconcile exact receipts

All 120 accepted turns record capability version, revision transition, tick window, typed details, and before/after fact hashes. A stale apply is rejected before mutation.

action → receipt → consequence
04

Publish, never operate

This browser fetches static JSON with GET only. Local files stay in memory and are never uploaded. Semantic validation happens before publication through the strict CLI replay.

dwarf-society replay

Canonical evidence · immutable replay

Founding Stability, tick by tick.

Loading reviewed publication…

Native observation plane

Fortress viewport

Frame
80 × 25
Glyphs
2,000 B
Boundary
Paused
Loading reviewed evidence…
Awaiting canonical replay. REV —

Typed management ledger

Observed job → priority receipt → 120 ticks

4 receipts
  1. Loading typed receipts.

Select a management receipt to inspect its fences and consequence.

Fail-closed proof

Stale state did not mutate

LOADING
Status
Loading rejection evidence

The rejected attempt remains visible as invalid_action_count=1; it is not erased to improve the result.

Immutable source · disposable episode

Cleanup is part of the score.

The admitted source checkpoint is mounted read-only, copied for the episode, then re-hashed after cleanup. A run cannot pass with residue or a changed source digest.

Owned processesLoading proof
Disposable installLoading proof
Disposable saveLoading proof
Private homeLoading proof
Install treeLoading immutable hash
Source save treeLoading immutable hash

Publication-time validated snapshot

One canonical result, not a public leaderboard.

SNAPSHOT V2

The browser sorts no mutable server state. This reviewed snapshot contains the deterministic reference-controller proof only; it does not claim model comparisons that have not been run.

ControllerResultFoundersFoodDrinkObjective / horizonEvidence
Loading reviewed snapshot…

What this result means

Validated result pending.

Demonstrated

  • Waiting for the reviewed publication manifest and evidence hashes.

Not demonstrated

  • Open-ended planning, conversation, civic governance, or individual dwarf agency.
  • Worker assignment, job creation, inventory mutation, or arbitrary DFHack control.
  • Generalization to another fortress, runtime pair, source case, or later curriculum stage.
  • A public evaluation service, model ranking, or live game operation.

Run the offline evaluator

Native execution. Strict replay. Separate from this site.

Stage identity and prerequisites come from versioned curriculum data. The runtime requires admitted local assets; the public site distributes no game binaries or saves.

Inspect curriculum
dwarf-society stage list
Apple Silicon proof entrypoint
scripts/run-stage1-apple.sh
Native x86_64 Linux run
dwarf-society stage run --stage 1 \
  --install-root /assets/install \
  --golden-save /assets/founding-stability-v2 \
  --state-root /state \
  --df-archive /assets/df_47_05_linux.tar.bz2 \
  --dfhack-archive /assets/dfhack-0.47.05-r8-Linux-64bit-gcc-7.tar.bz2
Strict offline evidence replay
dwarf-society replay trajectory.json

Prospective curriculum · manifest v2

Stable identities; future order changes prospectively.

  1. 0
    Terrain SurveyPrerequisite boundary proof
  2. 1
    Founding StabilityReal v2 proof · published here

There are no later stages in the published curriculum. New stages, order, or prerequisites require a new prospective curriculum version; they do not rewrite this scenario or result.