01 · Problem
What needed solving
The same file extension can describe very different kinds of 3D evidence. Calling every Gaussian PLY a learned reconstruction hides the difference between initialisation and native model output.
02 · Approach
How we're building it
We added a separate DA3 Nested path that exports native means, scales, rotations, opacities and colour coefficients, applies a deterministic cap, and records the source in the run manifest.
03 · Finding
What we learned
A verified export path is meaningful progress, but it is not the same as proving optimisation quality, metric scale or unseen-side accuracy. Precise provenance makes that distinction visible.
04 · Next question
What we are testing next
How should native Gaussian exports be evaluated against held-out views and real capture sequences before quality comparisons are published?
05 · Timeline
Where it is
- 2026Native DA3 inference completed on the public RTX PRO 6000 GPU
- 2026Verified one-view run produced 63,504 candidates and exported 2,048
- NextAdd held-out view evaluation and a named comparison baseline
Related in Spatial Intelligence