Product / Evaluations
Agents make assumptions when context is missing. Those decisions are checked against policy, and patterns of incorrect behavior surface across runs.
A threshold is exceeded. A deadline passes. A boundary moves. Each decision is checked against the conditions active at the time.
Evaluations
An agent fills a gap with an assumed value and moves on. That assumption is checked against the policy that should have applied, and repeat offenders surface across runs.
Agent decisions diverge from what policy intended. Decisions are tracked across runs, so drift shows up as a trend while it's still small.
New outputs need checking before they're trusted. The same policies that guide decisions run as checks against output, flagging anything outside policy before downstream use.