Silent Failures in Single-Shot Text-to-SQL: A Blind, Nine-Model Cross-Comparison on the *Ground Truth* Benchmark

We report a blind, single-shot text-to-SQL evaluation of nine large language models against a single frozen instance of the Ground Truth benchmark — a text-to-SQL and analytical-disclosure benchmark built on a synthetic retail warehouse seeded with eleven documented defect classes (fan traps, semi-additive sums, many-to-many overlaps, coverage gaps, non-response, staleness, empty buckets, and others). Each model received identical inputs (question text, prose schema documentation, and a column/format reference), produced one SQL SELECT and a structured disclosure per question with no execution feedback, and was scored by the benchmark's own deterministic-first grader. Headline composite trust scores span a narrow 0.10 band (0.704–0.804) across all nine models. Against that narrow spread, three structural findings hold: (1) a subset of failure families — average-satisfaction non-response and a lowest-sales empty-bucket trap — score identically or near-identically across every model tested, regardless of vendor or scale; (2) five different models routed through the same coding-agent host converge on nearly identical scores and the same specific wrong answers, while the same model run two different ways diverges from itself more than the nine models diverge from each other; and (3) roughly half of every model's scored "silent failures" are disclosure-taxonomy mismatches rather than arithmetic errors — the model names the correct hazard in prose but files it under a category the deterministic grader does not credit as equivalent. This deposit includes the full report (PDF and Markdown source) and the complete cross-model test kit: blind analyst instructions, the 111-question set, schema documentation, all nine raw per-model answer sets, the scoring driver, and the derived comparison tables and chart. No ground truth, acceptance sets, signatures, or key facts are included in the blind kit.