6 compilation targets9 shipped levels114 assertions1,954 compiled solids, 0 defects355k navmesh samples, 0 islands

AI should design a map program, not type its triangles.

A reproducible prototype for compiling semantic tactical-FPS layouts into brushes, CSG, exact CAD/B-rep, voxel/SDF, GLB/OBJ/STL, DXF, and engine-native primitives—with geometry and gameplay validation.

This is now implemented, and you can walk through it

The proposal below has been built: a LevelSpec schema, an eleven-pass deterministic compiler, geometry / movement / tactical validators, six exporters, and a first-person viewer. Six levels ship with it. Open the viewer → · README · the original contract

Measured by that implementation, on the levels it ships — every figure below this box is from the original study and describes different benchmark levels.

Level	Solids	Compiled: coplanar / overlapping	Naive build of the identical spec	Navmesh baked from the geometry
Compiler Demo Hall	175	0 / 0	214 / 214	100% reachable
Dust II (approximate)	300	0 / 0	344 / 538	100% reachable
Cache (approximate)	237	0 / 0	302 / 532	100% reachable
Clubhouse (approximate)	386	0 / 0	446 / 446	100% reachable
Ring Arena (original)	124	0 / 0	190 / 190	100% reachable
Three-Lane Two-Site	179	0 / 0	217 / 256	100% reachable
Three-Storey Siege	251	0 / 0	321 / 321	100%, 44% sealed
Vertical Stack Tower	302	0 / 0	163 / 163	100% reachable
Broken By Design	75	fails 11 gates, as intended	63 / 63	23%, 5 islands

In every naive build each box is individually watertight, consistently wound, and contains its own centroid as a six-plane brush. The report panel in the viewer states this explicitly next to the failure counts.

Compiled floor plans

Drawn by the compiler from the same boundary records as the 3D geometry, so the plan cannot disagree with the level you are standing in.

Cache — approximate reconstruction

The least precise of the three reconstructions: the area set and adjacencies are right, several finer connections are a reasonable guess.

Dust II — approximate reconstruction

A topology reconstruction with the real callouts, not a measured copy. Accuracy needs an authoritative source as the LevelSpec input.

Ground plan of the Dust II reconstruction, all named areas labelled

Clubhouse — approximate reconstruction

Basement, 1F and 2F on an aligned footprint, with soft walls, reinforcement slots and hatches as semantic roles.

Three storey plans of the Clubhouse reconstruction

Three-Storey Siege Benchmark

Three storey plans: basement, ground floor with courtyard, first floor

Three-Lane Two-Site Benchmark

Ground plan with three lanes and two sites, plus the catwalk layer

Vertical Stack Tower

Ring Arena (original)

Ring plan with a sealed centre and an overwatch balcony

Central result — original study

Both maps are made from individually watertight, consistently wound boxes. Yet the Dust-style collection still contains 156 coplanar face pairs and 136 intersecting-volume pairs; the R6-style collection contains 304 and 259. Primitive validity does not imply collection validity. A deterministic ownership and Boolean/bake step is still necessary.

Reference topology studies from the original study

CS-style two-site layout

Approximate three-lane topology with an explicit elevated Catwalk layer and two stair links.

Labeled floorplan of the Dust-style topology benchmark

R6-style three-storey layout

Approximate legacy Clubhouse study with rooms, stairs, hatches, soft/reinforced walls, windows, and rappel entries.

Labeled floorplans of the three-storey R6-style topology benchmark

Representation comparison

Path	Dust-style	R6-style	Use
Separate primitives	122 boxes / 1,464 tris	229 boxes / 2,748 tris	Editable, not runtime-ready until internal/coplanar faces are resolved.
Quake brushes	122/122 half-space tests pass	229/229 pass	Excellent deterministic greybox format.
OpenSCAD union	1,414 tris; watertight	2,396 tris; watertight	Offline CSG bake.
CadQuery B-rep	Valid STEP; 1,414-tri tessellation	Valid STEP; 2,396-tri tessellation	Exact solids, curves, parametric CAD, validation.
Voxel/SDF	162,908 tris; watertight	188,840 tris; watertight	Robust repair/destruction fallback; high cost and quantization.
Gameplay graph	2,233 nodes; 1 component	1,342 nodes; 1 component	Connectivity and route metrics independent of rendering.

Recommended architecture

brief / plan / reference
        ↓
AI edits a constrained LevelSpec
        ↓
schema + topology + solver checks
        ↓
deterministic compiler owns shared walls, openings, winding and axes
        ↓
static shell → robust brush/CSG/B-rep bake
runtime surfaces → separate soft walls, hatches, windows and rappel actors
        ↓
collision + nav + LOS + route timing + simulated playtests
        ↓
engine import and designer review

For exact private reconstruction, use an original/authorized source, official blueprint, or a validated decompile as the authority. For generation, create a new LevelSpec under the same route, timing, verticality, sightline, and destruction constraints.

Read and run

Research report

Failure taxonomy, survey of brushes/CSG/CAD/BIM/SDF/DSL/WFC/quality diversity/world models, and experiment results.

Open research_report.md

Implementation blueprint

A staged product design for Three.js, Unity, Godot, and Unreal.

Open implementation_blueprint.md

LevelSpec contract

The AI-facing schema, invariants, portals, layers, wall roles, and vertical connections.

Open LEVELSPEC.md
What the implementation added (1.1)

Rebuild and test

Run npm run compile to build and validate every level, and npm run check to typecheck, test and compile.

Open the viewer