Does the newer YOLO26-pose model separate stacked stripes better?

BLOCKED: a GPU pod was allocated but preflight checks failed before training could start — a data-integrity issue (unreadable retained image, unresolved EXIF frame provenance), not a compute-availability problem.

What partial work found

The NMS-free hypothesis remains untested. The current input snapshot contains a test fixture instead of a photo and unresolved image/label frame conflicts. Establish the correct frames and source image before measuring the finder swap; otherwise a recall difference could be a data error.

Check	Result
GPU pod allocated	ubgu8igk4t349a (NVIDIA RTX A5000)
Training started	No
Why stopped	CPU preflight failed before training: unreadable retained image and unresolved EXIF frame provenance. This was not a RunPod capacity/authentication failure. Pod was still initializing when terminated.
Unresolved image/label frame conflicts	3 photos

Source: model-yolo26-swap.json, pre-relabel baseline attempt.

raw data (JSON) →