MVP-v1.0 · evidence-first spatial reasoning

See where 3D question answering fails—and why.

A self-running, captioned walkthrough of a queryable spatial-graph reasoner over real captured scenes. It isolates box construction, learned instance extraction, and learned labels while keeping every answer tied to visible 3D evidence.

Launch the two-minute demo View source and results Open the legacy toy demo

Real captured geometry

Orbit full-resolution colored Replica meshes, then compare oracle and learned instances.

Stage-isolated failures

Watch an office answer collapse at instance extraction and a room_2 support anchor disappear under one label error.

Honest uncertainty

Human keys, answer/empty/defer/unknown states, disclosures, and click-level provenance remain visible.

Scope: C1 and C2 are declared evaluation-only ladder stages. This demonstrates attributable structural reasoning over a frozen Replica reconstruction; it is not an end-to-end NeRF/3DGS claim.