| Generalize from a slow implementation | Another algorithm may avoid its work. | Prove a requirement for every admitted transcript or procedure. |
| Count distinct inputs without considering acceptable answers | Several inputs may share one output. | Count output-distinct classes. |
| Supply an inconsistent adversary | No admitted input could produce its replies. | Retain a nonempty compatible input family. |
| Fix randomness, then choose a different hard input for each seed | This can miss a successful randomized algorithm. | Use one prior input distribution or a valid coupled argument. |
| Ignore input-dependent preprocessing or stronger access | The algorithm already acquired the supposedly missing information. | Charge or expose that resource and restate the model. |