A drawbridge office adapts a general model to bridge-status phrasing and then finds it can no longer answer the everyday weather questions it used to handle. What experimental metric is missing?
Select an answer to reveal the explanation.
Short Explanation
The model learned bridge-status phrasing and forgot everyday weather. Score a holdout from the old skill alongside the new task so forgetting is visible. A new-task win alone, a memory leak, or a storyboard will not show that cost.
Full Explanation
A transfer win on the new task can hide catastrophic forgetting of a skill the office still needs. The missing readout is a holdout from the retained old task, scored with the new-task set after adaptation. If weather answers collapse while bridge-status improves, the experiment has a cost that a single new-task number conceals. Memory leaks, skipping the old skill, and multimodal storyboards are not that holdout.