Skip to main content

Built with AILANG

Godot 4 renders the game. An AILANG simulation runs as a child process and talks newline-delimited JSON over stdin and stdout. The decision and its reasons are in ADR 0001.

┌─────────────────────────────┐   one NDJSON line per tick   ┌──────────────────────────────┐
│ Godot 4.7 (Forward+, Metal) │ ───── input: intents ──────▶ │ AILANG simulation            │
│  starfield + sky shaders    │                              │  sim/ship.ail   I/O shell    │
│  galaxy map, HUD, input     │ ◀──── state: changes ─────── │  sim/core.ail   pure core    │
│  draws what the sim says    │                              │  sunholo/relativity (pinned) │
└─────────────────────────────┘                              └──────────────────────────────┘

The simulation​

  • The world state lives in AILANG. Ship, the two clocks, journeys, the commit rule, the energy ledger and the random streams are all simulation state. Godot sends player intents and draws the state it gets back.
  • A pure core on the bytecode VM. sim/core.ail passes --strict-bytecode: it runs entirely on the AILANG VM with no interpreter fallback. I/O lives only in the thin sim/ship.ail shell.
  • Physics from one package. The formulas come from sunholo/relativity, a published AILANG package with its own tests and check values. The simulation pins a version; GDScript and the shaders mirror it.
  • The rules are the simulation's. A committed journey cannot be cancelled because the simulation refuses the intent, not because a button is greyed out.

Determinism​

The simulation has no hidden state. Given the same seed and inputs it produces byte-identical output.

  • Randomness is SplitMix64 with six named streams, checked against published vectors.
  • make replay re-runs recorded input logs and diffs the state. A 10,000-tick session is byte-identical on the VM and the interpreter, and matches its recorded golden.
  • When the VM and the interpreter disagree, that is an AILANG bug: it gets shrunk to a minimal repro and reported upstream. The game is meant to stress-test the VM: thirteen AILANG issues were filed or tracked during the journey-core milestone alone.

A real exchange​

Trimmed from tests/replays/alpha_cen.ndjson and its recorded state log (… marks elided fields). Godot plans a trip to α Centauri at rapidity 2.6466 (0.99c), and the simulation answers with the plan:

{"v":2,"type":"input","tick":1,"dtau":0,"intents":[{"k":"plan","target":{"index":1,"id":"alpha Cen","pos":{"x":0,"y":0,"z":-4.37}},"cruise_phi":2.6466524123622457}]}
{"v":2,"type":"state","tick":1,"status":"ok","changes":{"journey":{"state":"planned","plan_id":1,"plan":{"cruise_beta":0.99,"cruise_one_minus_beta":0.01000000000000001,"cruise_gamma":7.088812050083356,"ship_years":0.6226958707592057,"earth_years":4.414143655383168, …}}}}

Note cruise_one_minus_beta: near c the simulation carries 1 − β itself instead of letting anyone subtract.

Pipelines in AILANG too​

The offline tools are AILANG, with Godot headless doing image I/O:

  • the catalogue builder turns CNS5, Gaia GCNS and Hipparcos into binary tiers (the 331,312-row large tier builds in under a minute on the VM);
  • the destar pass removes catalogue stars from the Milky Way panorama;
  • the sky-model fitter gives each panorama texel a colour temperature on the Planckian locus.

Python is allowed only as an independent oracle, a test harness or a throwaway spike, and a guard in make test enforces it.

Tested before it merges​

make test runs headless in CI: the physics reference checks, the simulation against closed forms, VM/interpreter parity, the strict-VM core and the replay goldens. make golden checks the GPU shaders against the CPU reference to sub-pixel accuracy, and make capture renders the reference images that get looked at by eye.