ROUTING METHOD · QUALITY EQUIVALENCE UNMEASURED
Cheaper-Model-First Routing Needs a Quality Gate
A long-scroll visual companion to the article. It summarizes the article evidence boundary and method; it does not extend the underlying result.
At a glance
Evidence in scope
- Keep task and verifier fixed
- Record retries and usage separately
- Escalate when the rubric fails
Evidence boundary: No current matched model-price/quality benchmark supports a cheaper-model winner or savings estimate.
How to use the diagram
Preregister representative tasks, model versions, rates, retries, and a blinded acceptance rubric before comparison.
What this visual explains: A lower-priced route is only a candidate until it passes the same task-specific acceptance bar.
What it does not establish
No savings, quality equivalence, or winner is measured; refresh official rates before a dated estimate.