Error Detection Rate
Catching contradictions and bad inputs.
What it measures
Catching contradictions and bad inputs.
Where it sits
Error Detection Rate is one of the Model pillar’s dimensions (10%of the composite). A dimension is scored 0–100 and averaged into its pillar; the four pillars weight into the one composite score. See Pillars and weights.
How it’s graded
Objective — scored deterministically from what the agent actually did on the probe, with no judgement. Every score comes from a test that actually ran — no self-report.
Improving it
This measures the base model's raw capability, not the harness you build on top — it moves when you change model, not when you change your scaffolding.