| Useful coverage and cheap proxies | Test the consequential distinction rather than the easiest correlated signal. |
| Improvement speed and qualification cost | Start with a discriminating counterexample; expand only for the intended reliance. |
| Human judgement and executable checks | Use the means capable of observing the relevant property; neither is universally superior. |
| Aggregate evaluation and individual failures | Inspect the cases that could defeat the intended use. |
| Stable reuse and changing models or tools | Bound the qualification and identify what would reopen it. |