Learning goal: Learning curves vary sample count and probe the potential benefit of more data; capacity curves vary model size; training curves vary optimization time. Interpret each with uncertainty rather than treating its shape as a unique diagnosis.

Training and validation errors versus n; bands: repeated-run interquartile ranges; horizontal reference: σ².

Second plot: capacity versus degree or training progress versus steps.

Observe: changing d from 2 to 12 can lower the learning-curve plateau and change the useful next comparison. A smaller validation set broadens evaluation uncertainty. The training view can show decreasing training error alongside nonmonotone population risk; the useful stopping region may be broad in this toy. Its width is not a universal early-stopping guarantee.