Enterprise SQL agents hit a 99% silent-divergence risk

ESQ-Bench tested natural-language-to-SQL systems across 465 tables, 164,682 rows, and 550 gold-validated question-query pairs, finding operational silent divergence in 73% to 99% of queries that executed successfully. The benchmark paper says Claude Sonnet 4.6 still fell from 87.4% to 68.7% execution match across its three complexity tiers, so data teams should validate result semantics—not stop at a green query—alongside the archive’s enterprise harness analysis.