Atlas World generates deterministic Blender observations, compiles each RGB image through Atlas, and measures the prediction against isolated renderer truth. The compiler is handed the image and nothing else — no truth camera, no scene spec. If the solve is right, it earned it blind.
A benchmark is only honest if the thing being tested can't peek. Atlas World enforces that at the boundary: the compiler receives four public inputs and is structurally denied everything else.
Each WORLD_000001 directory holds an observation, Blender truth, Atlas prediction artifacts, evaluation metrics, and an atomically updated manifest.
These are the real evaluation keys the harness computes per episode, then aggregates by median. The values are blank on purpose — this page ships no results until a run produces them.
Status awaiting run — missing optional depth or geometry stays explicitly not_available, never guessed.