# Failure modes

- calling attribution causal
- holdouts with contamination
- post-hoc metric switching
- Hiding uncertainty behind confident language
- Treating platform attribution or generated text as causal/customer proof without independent support