Results

Leaderboard

Every routing policy of the recorded run, ranked by quality for each agent, with what it spent.

The routed policy sends each question to the cheapest tier that can settle it. The board shows what that costs in quality: on matching it trails only the policy that pays the API model for everything, and spends less. Other policies are the alternatives it was compared with: local model only, API model only, evidence only, and routing under a cap on API calls.

How to read it

• Quality is the headline metric of each agent: F1 after repair for matching, precision for extraction, and correct placements for reasoning.
• API spend is list price, and each policy is charged only for the questions it asked. Local and evidence-only policies spend nothing.
• Click a column to sort. Ranks stay with quality, so a sort does not change who is first.

The showcase shows every decision behind these numbers.