Quality-constrained model routing

Find the cheapest route that still clears the quality floor.

Evaluate candidate model routes across a fixed task set, reject any route below the quality threshold, and explain the lowest-cost passing combination.

Try the fixture No sign-in · precomputed from the tested fixture

One product.

One decision.

Visible evidence.

Interactive proof

Reveal the verified fixture.

This exact 3-model × 20-case CSV is evaluated by product.py.

Routing experiment

Quality floor: 91

Evaluation results:
case_id,task,model,quality,cost
C01,extract,gpt-5.6,96,2.10
C01,extract,terra,92,1.00
C01,extract,luna,90,0.40
C02,reason,gpt-5.6,96,3.20
C02,reason,terra,92,1.50
C02,reason,luna,85,0.70
C03,draft,gpt-5.6,93,2.40
C03,draft,terra,92,1.10
C03,draft,luna,89,0.50
C04,extract,gpt-5.6,97,2.10
C04,extract,terra,93,1.00
C04,extract,luna,91,0.40
C05,reason,gpt-5.6,95,3.20
C05,reason,terra,91,1.50
C05,reason,luna,84,0.70
C06,draft,gpt-5.6,94,2.40
C06,draft,terra,93,1.10
C06,draft,luna,90,0.50
C07,extract,gpt-5.6,96,2.10
C07,extract,terra,92,1.00
C07,extract,luna,90,0.40
C08,reason,gpt-5.6,96,3.20
C08,reason,terra,92,1.50
C08,reason,luna,85,0.70
C09,draft,gpt-5.6,93,2.40
C09,draft,terra,92,1.10
C09,draft,luna,89,0.50
C10,extract,gpt-5.6,97,2.10
C10,extract,terra,93,1.00
C10,extract,luna,91,0.40
C11,reason,gpt-5.6,95,3.20
C11,reason,terra,91,1.50
C11,reason,luna,84,0.70
C12,draft,gpt-5.6,94,2.40
C12,draft,terra,93,1.10
C12,draft,luna,90,0.50
C13,extract,gpt-5.6,96,2.10
C13,extract,terra,92,1.00
C13,extract,luna,90,0.40
C14,reason,gpt-5.6,96,3.20
C14,reason,terra,92,1.50
C14,reason,luna,85,0.70
C15,draft,gpt-5.6,93,2.40
C15,draft,terra,92,1.10
C15,draft,luna,89,0.50
C16,extract,gpt-5.6,97,2.10
C16,extract,terra,93,1.00
C16,extract,luna,91,0.40
C17,reason,gpt-5.6,95,3.20
C17,reason,terra,91,1.50
C17,reason,luna,84,0.70
C18,draft,gpt-5.6,94,2.40
C18,draft,terra,93,1.10
C18,draft,luna,90,0.50
C19,extract,gpt-5.6,96,2.10
C19,extract,terra,92,1.00
C19,extract,luna,90,0.40
C20,reason,gpt-5.6,96,3.20
C20,reason,terra,92,1.50
C20,reason,luna,85,0.70

The tested result is ready to inspect.

Reveal the expected decision generated by the product's deterministic acceptance fixture.

How it works

Small surface. Clear chain of proof.

Fix the eval

Hold cases, scoring, and the quality floor constant across candidates.

Enumerate

Search complete routing combinations instead of hand-picking a winner.

Select

Choose the lowest measured cost among routes that pass every constraint.

✓ Exactly 20 single-task cases✓ Quality bounded to 0–100✓ Positive baseline cost