6.8: Understanding Layer Produces Net Harm

If the understanding layer's error rate on value attribution produces worse outcomes than a no-understanding baseline, the layer is net harmful and must be constrained or removed.

Distinct from 6.2 (AI plateau, which asks whether the capability exists). This asks whether the capability, even when present, is safe to deploy. A system that confidently misattributes values and then uses that misattribution to time observations during degraded states could actively harm the person it is designed to serve.

E3 must test not just whether the understanding layer adds value, but whether it adds value safely -- specifically, whether value attribution errors during degraded states produce worse outcomes than the no-understanding control.