Cannot reproduce Table 2 LingBot-World scores

[theREALevan]

Hi, thank you for releasing MBench.

We are evaluating the 227 released LingBot-World items with
subset = environment in
MBench-A/models/lingbot_world/samples.jsonl, using the public code at
commit c44dbd9.

Using the released videos, DA3-GIANT-1.1 over all frames, and the public
mbench eval pipeline, we obtain:

Metric	Table 2	Ours
Epipolar	22.12	0.13
Reprojection	7.57	1.14
Lighting	40.06	89.62
Style	85.87	87.06

All samples pass mbench validate. The rollout videos are byte-identical
to the public release.

Could you please share the exact DA3 artifact generation setup used for
Table 2 (checkpoint, preprocessing, frame sampling, pose convention), or
release the LingBot-World results.npz and corresponding input_images?

Also, could you confirm whether Table 2 used the current public evaluator
and the same released environment subset?

Thank you!