hidden-state monitoring for safer dialogue

Some harmful conversations do not look obviously harmful at the surface level. A reply can seem normal while the conversation is gradually moving toward a risky goal.

That is the motivation behind AERIC. Instead of only asking whether the current text is unsafe, the project looks at whether internal model behavior changes as a conversation develops. The research question is simple: can hidden-state shifts help detect implicit harmfulness earlier than output-only classifiers?

The current direction is to compare conversation-level risk signals across turns, then check whether those signals align with human judgments of escalation.