Would a centerline-aware training loss finish faint stripes more completely?

BLOCKED: GPU access failed (RunPod returned 403 on every credential source tried); only CPU diagnostics ran, and the underlying question is also gated on whether SAM3 stays in production at all.

What partial CPU work found

No GPU training or inference ran. This test is also conditional: it only matters if SAM3 stays in the production pipeline, and undecided.

Check	Result
GPU training ran	No
Paired holdout photos evaluated	0
SAM3 production decision	undecided

Source: sam3-cldice-loss.json.

raw data (JSON) →