We don't claim to be "good" in the abstract. We take the same model, answering the same database question, two ways: with SIXTA available and without it. The difference is what SIXTA adds. We run the whole suite several times and report the average, because a single model run is noisy.
The assistant answers from its own knowledge. The baseline.
Same model, same question, with the SIXTA tools available.
The deterministic tool output, no model. The ground reference.
129 cases across MySQL and PostgreSQL and all 13 tools, graded blind (our branding stripped) by a majority vote of three independent models. Each metric is computed per run; we report the mean and its range. The connected arm holds steady run to run; the bare-model baseline bounces, so we average it and show the spread rather than cite one run.
| Model | Critical-issue recall | Unsafe migrations approved | Flatly wrong answers |
|---|---|---|---|
| Haiku (small) | 75% → 97% | ~7% → 0% | ~14 → ~1 |
| Sonnet (mid) | 90% → 100% | ~5% → 0% | ~4 → 0 |
| Opus (frontier) | 94% → 100% | ~5% → 0% | ~5 → 0 |
On the recommended tiers (Sonnet, Opus), connected answers are flatly wrong zero times, every run. Haiku is the budget tier: connecting cuts its wrong answers from ~14 to about 1 (it still slips occasionally on the hardest version-specific cases), so reach for Sonnet or Opus when correctness is critical. Connected safety is stable across every run: 0% unsafe migrations was 0 of 75 trial runs for every tier. "Unsafe migrations approved" = the model called a table-locking or data-rewriting migration safe, the most consequential error.
No connection to set up. The free Connect tier works only from the SQL you paste: no credentials, no access to your systems. Connecting to your live database is the paid Connect Pro tier (on demand) or the SIXTA platform (continuous).
Nothing stored. Pasted SQL is analyzed in memory and never persisted.
Deterministic verdicts. Same artifact, same analysis, from version-aware rules. The model relays them, it doesn't invent them.
Self-checked fixes. Every rewrite is re-analyzed before you see it, and flagged if it isn't clean.
129 cases · MySQL & PostgreSQL · 3-model blind judge panel · averaged over repeated runs · last updated 20 Jun 2026. Architectural items above are guarantees; the figures are measured results on our own suite, not absolute guarantees for every possible input.