Queries in plain language, and where that stops
Asking a question in English and getting a chart is a good default and a bad contract. What we let it do, what we make it show you, and the line it does not cross.
Typing "how many people who signed up in June came back in July" and getting a chart is a genuinely better default than learning our query syntax. It is also a bad contract if the tool stops there, because you now have a number whose derivation you cannot inspect.
Show the query, always
Every plain-language question renders the query it produced, above the result, in a panel you cannot collapse away. Not in a menu, not behind a toggle that remembers being closed. If the interpretation is wrong, the wrongness is usually obvious in the query and invisible in the chart.
The common failure is not a broken query. It is a reasonable query answering a slightly different question than the one you asked, most often about which window "came back" refers to. You will catch that in two seconds reading the generated filter and never by looking at a line going up.
Where it stops
We do not let a generated query define a metric that other things depend on. Ad-hoc exploration, yes. The definition of activation that six dashboards and a weekly report are built on, no. That gets written once, by a person, with a name and a comment, and it changes through a review.
The distinction is not about trusting the generation. It is that a definition everyone depends on should be legible in one place, and a thing that regenerates from a sentence is legible in as many ways as the sentence can be read.
The honest limit
It is good at aggregation and filtering, weaker at anything involving ordering across sessions, and it will occasionally produce something confidently wrong. That is why the query panel does not collapse. The feature is not the natural language. The feature is that you can always see what it did.