Current AI safety evaluations measure whether individual responses to children avoid harmful content. But developmental risk doesn't emerge from single moments—it accumulates through patterns of interaction over time.

Across five scenarios evaluated on frontier models, we found that systems can be fully policy-compliant, empathetic, and content-safe while still generating higher-severity developmental risk through repeated validation without containment, relational substitution, and caregiver displacement.

All cases passed conventional turn-level review; severity divergence appeared only through longitudinal evaluation. Read the full Cross-System Evaluation →

If AI systems are to engage children safely, trajectory-level interaction patterns must become a first-class safety object. The attached brief documents findings from a structured cross-system evaluation of five frontier models across developmental risk scenarios — run under the ANCHR_v1 framework.

See what an evaluation produces. Review a sample audit report →