Deletion Test
A method for judging whether a rule in CLAUDE.md is necessary: after a given instruction is deleted, if the AI makes no error under a fresh prompt, that instruction should be removed. The method turns rule effectiveness into a verifiable experiment.
Timeline (last 90 days)
Deletion test: if deleting a given instruction would not cause the AI to make mistakes under a fresh prompt, that instruction should be deleted.
The coupling between detection strength and faithfulness is not a defect of a specific explanation method, but is more likely related to the deletion-based evaluation paradigm itself.
Deletion testing evaluates explanation faithfulness by masking the regions highlighted by the explanation and observing the degree of drop in model confidence.
All Facts (3)
Deletion test: if deleting a given instruction would not cause the AI to make mistakes under a fresh prompt, that instruction should be deleted.
50%UnverifiedThe coupling between detection strength and faithfulness is not a defect of a specific explanation method, but is more likely related to the deletion-based evaluation paradigm itself.
50%UnverifiedDeletion testing evaluates explanation faithfulness by masking the regions highlighted by the explanation and observing the degree of drop in model confidence.
50%