How it works

We learned this the hard way

Until 13 October 2026 this site's SQL page had charts that each held their own SQL, written by hand over the tables. Every rule the model learned had to be copied into it, and it fell behind: 84 of 683 checked questions gave a different number (trailing spaces in text, how a fixed decimal average rounds, the type of an IF). Since then both pages ask the model, and 0 of 683 differ.

My take (a personal opinion)

• With AI in the loop it gets worse: an agent fixes a wrong number where it sees it, in the query, so the rule is copied once more, a little differently each time.
• A small surface means fewer errors. A visual here can only name fields and filters; a lint refuses arithmetic in the page.
• The cost: a new figure is a new measure in the model, not a quick line of SQL in the chart.
• SQL in the front end still has its place for exploration: the dashboard's Analyze tab takes any SQL.