Quality gate

Savings that come from worse hard answers are a product failure.

Savings without a quality score are marketing. Promptimizer measures both.

Two checks

Online: after a cheap completion, empty, too-thin L4+, or refusal answers escalate. Benchmark: each of the 15 gold tasks is scored with required-concept coverage + structure, routed and frontier.

What degraded means

Missing required concepts, a thin outline on system design, or an explicit small-model refusal. Short L1 factual answers are allowed.

How to fail on purpose

Route everything to nano. saved_pct looks heroic. avg_quality_routed collapses on the five hard rows.