· Perseval · Product principles
Exact Groups First, Optional AI Second
Keep failure identity deterministic, use similarity only for navigation, and treat hosted judgments as unproven assistance.
Keep failure identity deterministic, use similarity only for navigation, and treat hosted judgments as unproven assistance.
A practical loop for turning recurring agent failures into evidence-backed eval definitions.
Outcome-only scoring misses the decisions, tool calls, and recoveries that make agent behavior reliable or fragile.