Recipe: Cost-aware evaluation

Recipe — score cost alongside quality so you can pick smaller-and-cheaper when it's good enough.

Capture cost per row from traces. Compute a composite metric: quality / cost.

def cost_quality(scores, costs):
 return [s / max(c, 1e-6) for s, c in zip(scores, costs)]

Use this to spot when a smaller, cheaper model wins on the cost-quality frontier.

See also

• Concept: ECU credits