Judgment Benchmarking

Judgment Benchmarking turns raw decision telemetry into comparative context. Instead of only showing your own override rate or escalation rate, DelegateZero shows how those numbers stack up against the broader network so you can see where your operating model is unusually weak, unusually conservative, or unusually strong.

There are two surfaces:

• Judgment Health dashboard at /app/benchmarks for workspace-specific comparisons and recommendations
• Public benchmarks page at /benchmarks for anonymized aggregate benchmark data

What is benchmarked

The initial benchmark set focuses on the signals DelegateZero already records with high confidence:

• Override rate by decision category - how often users correct the system in categories like refunds, approvals, or routing
• Escalation rate by cohort - how often decisions still need human review
• Average confidence by tenure milestone - how confidence changes across the first 30, 60, and 90+ days of usage
• Context depth at first successful autonomous week - how much context workspaces had loaded when they first completed a full week with zero escalations

Additional slices like ARR band and company stage can be layered in once that segmentation data is collected explicitly.

Methodology

Benchmarks are computed from anonymized, aggregate telemetry. No benchmark surface reveals raw workspace content, individual requests, policy text, or customer data.

The current implementation uses these working definitions:

• Override rate - percentage of decisions in a category that were later marked as overridden
• Escalation rate - percentage of completed decisions whose final decision was escalate
• Confidence milestones - confidence scores grouped by how many days had elapsed since a workspace was created when the decision happened
• First successful autonomous week - the first calendar week with at least 5 completed decisions and zero escalations

How to use it

Use Judgment Health as a prioritization system. If your category-specific override rate is above benchmark, tighten the policy or add better precedents. If your escalation rate is high but your confidence is already near cohort average, the issue is usually policy coverage or threshold tuning rather than model uncertainty alone.

The point of benchmarking is not vanity. It is to tell you where your judgment system is under-specified relative to peers who are already operating more autonomously.