Repository navigation
Vision & pillars
Condensed from docs/00-vision.md.
- Simulation ⊥ rendering. "Simulation is cheap, rendered simulation is expensive. Hence separating the two allows for depth and performance." Off-screen reality is lightweight state math; the GPU projects only the observed slice.
- Causality is an observation. An unobserved river is an equation in time. Arrive at tick 100 and the bed is dry, the trees are dead — the 50 intermediate steps never ran, and the player's mind supplies the story. Deterministic, so every visitor sees the same endpoint.
- The asset is the algorithm. No sculpting, no PNGs. A pig is a signed-distance function plus a color function; animation deforms space over time. This is what lets an all-LLM team have an art pipeline at all, and what makes a mod a few kilobytes of text instead of 50 MB of binaries.
- The world is a ledger. Actions are signed transactions; sleep/epoch boundaries bake them into a Merkle root. Two clients prove they agree by exchanging 32 bytes.
- Modding is the native interface. The engine is built from text-defined assets, compiled circuits and sandboxed WASM — so the public mod API is the internal API, not a bolt-on.
- One coordinate truth. Orientation, gravity and reference frames are specified before anything else is written, because everything else floats without them.
| Minecraft's lever | Ours |
|---|---|
| Sold unfinished alpha, updated fast on community feedback | Ship the Phase 1 engine as a usable platform early; players are co-developers |
| Let's Plays were free marketing because every playthrough was unique | A world that visibly ages between visits produces tellable stories for free |
| Zero-friction access (a Java applet in the browser) | WebGPU target: a browser tab is a first-class client |
| Java bytecode was accidentally decompilable → MCP, Forge, Fabric | Modding is deliberate: hot-reloaded text assets + WASM, no reverse engineering |
| Redstone attracted engineers | Machines that stay computable while unobserved — factories keep producing mathematically |
| Pocket Edition, consoles, then a C++ rewrite (Bedrock) for reach | Cross-platform from day one via WASM + WebGPU, not a future rewrite |
| Villager AI stayed simple because per-entity 3D pathfinding is expensive and builds need predictability | Aliveness comes from the macro layer (populations, water tables, growth), not more per-entity AI |
| MineColonies / Millénaire already do unloaded-town + realized-town | That trick, generalized, is the engine |
Not copied: tick-by-tick cellular redstone (incompatible with pillar 2), stored-per-block world saves, mesh/texture asset pipelines.
You leave a region for a week of in-game time. Return, and the river you damned is gone, the meadow past it has browned, the wolf pack has moved on, and the sawmill you built kept producing planks the whole time — rendered instantly, no chunk loading, no catch-up lag. The world aged while you weren't looking, and it can prove it aged the same way for everyone else.
The source conversation is an architecture dialogue. It contains no gameplay design, no art
direction, no UI, no lighting model, no save-file format, no minimum-spec performance budget
(it aspires to "60+ FPS" but names no GPU, no resolution and no per-pass frame allocation — that is
OQ-39) and no legal/licensing position — a gap the source genuinely left, whose licence half the
project closed on 2026-10-11 with MIT OR Apache-2.0 while product naming is still open
(OQ-43). The budget in docs/04 is ours, not the source's: geometry
≤ 8 ms + shading ≤ 4 ms + culling ≤ 1 ms + upload ≤ 1 ms = 14 ms committed, 2.7 ms headroom
against a 16.7 ms frame at 1080p/60 — provisional until spike S1 measures a real frame-time curve.
The gap items are logged as [gap] items rather than invented — see
Open questions. An engine is not a game; OQ-04 (what do you actually do?)
is a real risk to the project, not paperwork.