Does constraining the decoder actually help?

A head-to-head eval: GPT-5 with grammar-constrained decoding vs. without, on real ClickHouse.

Every query runs against the default.nyc_taxi table from ClickHouse’s public NYC Taxi dataset — 20M trips, 2015-07-01 to 2015-09-30.