Recount the same saved validation predictions at the user-selected per-axis threshold: 548/1104 = 49.64%. Retain the original 20/1000 and strict matching results, and label RL improvement as a future objective. Preserve weights, inference runtime and original evaluation records.
Add visual model card and five saved validation examples
Add a visual overview, English/Chinese cards, unchanged AgentNet input frames, exact saved step3368 predictions, and a text failure example. Preserve model weights and inference code.
