Sequential fine-tuning on tasks A, B, C is often summarised by final average accuracy. That single number can hide severe catastrophic forgetting: the model may do well on the newest task while having lost most of the earlier ones.

A more informative reporting format:
1) A task-by-checkpoint matrix: rows are evaluation times (after A, after B, after C), columns are task accuracies. Forgetting on task A = max earlier accuracy - final accuracy on A.
2) Intransigence / forward transfer: how well the model learns a new task compared with training on that task alone.
3) Backward transfer: whether learning a later task improves an earlier one.
4) Memory and compute overhead: replay buffer size, parameter growth per task, and training FLOPs. A method that stores all past data is not comparable with a fixed-memory method.

Controls that matter:
- Task boundaries: some methods assume known task identity at inference; state whether that is the case.
- Fixed vs growing capacity: adding a new adapter per task changes the comparison.
- Same data order? Run multiple task orderings, since order strongly affects forgetting.
- Multiple seeds with confidence intervals; continual-learning results are noisy.

A useful sanity check is a lower bound (fine-tune sequentially with no mitigation) and an upper bound (joint training on all tasks). Without both, it is hard to tell whether an improvement is meaningful.

Takeaway: report the matrix and the bounds, not just the final average. Forgetting is a per-task, time-dependent quantity.