When convergence curves lie

A convergence curve can create a false sense of certainty.

If the curve uses the wrong record, mixes current-generation values with best-so-far values, or hides infeasible solutions inside the plotted series, it may look like the algorithm is improving when the experiment is actually unstable.

What I check:

• Is the curve plotting current value or best-so-far value?
• Are infeasible solutions included?
• Is every point computed from the same objective function?
• Does the final point match the reported result table?
• Are different algorithms compared under identical run settings?
• Is the curve smoothed or post-processed?
• Does the curve reflect minimization or a transformed score?

A curve is not evidence by itself. It becomes evidence only when its recording logic is clear.