How RowFold checks its own numbers

Every figure an AI answer states is re-computed against your live data before you read it, and you can open the result to see exactly which query was replayed and what came back. This explains what that check does, how to read the badge, and — just as importantly — what it does not promise. Why an answer needs a receipt A language model can write a confident sentence containing a wrong number, and nothing about the sentence looks different when it does. That is the single reason not to trust AI with an operational question — so RowFold does not ask you to. Instead of promising the model is right, RowFold checks it. After the answer is written, every figure in it is re-computed against your live tables and compared to what was claimed. You see the outcome on the answer, and you can open the whole working. What actually happens While answering, the assistant reads your data with read-only tools — querying a table, aggregating a column, opening a record. When it states a figure that came straight from one of those calls, it also records which call produced it. The moment the answer is finished, RowFold replays each of those calls against your database as it stands right then, and checks whether the claimed figure is in the fresh result (allowing 0.5% for display rounding). Up to five figures are checked per answer. The replay is a read — it changes nothing, costs no AI usage, and the model has no part in it. Reading the badge The footer of every AI answer says which of three things happened: • ✓ N figures verified — every checked figure came back from your live data. • ⚠ N figures unconfirmed — at least one did not. The answer is still shown, because the reasoning may be sound and only one number stale, but treat those figures as unchecked and open the panel to see which. • no figures to check — the answer made no numeric claims. This is deliberately different from staying silent: nothing failed, there was simply nothing to re-run. Open it and read the working Click the badge and the receipt opens inside the conversation — your chat stays exactly where it is. For each figure it shows the claim, the tool that was re-run, the exact input it was re-run with, and the real result that came back, verbatim. The step chips above the answer work the same way. Each one names something the assistant did while investigating; click it to see the query it ran and the data it got. Nothing there is written by the model — it is the record of what was executed. Charts are checked too When an answer rests on a breakdown — visits by reason, revenue by month — a small chart appears under it. The assistant does not supply the bars; it names the aggregate, and RowFold re-runs that aggregate and draws whatever comes back. A chart therefore cannot disagree with your data even if the prose above it does. What verification does not promise It is a check on arithmetic against live data, not a judgement about meaning. Specifically: • Figures the assistant worked out itself — a percentage, a difference between two totals, a growth rate — are not re-run, because there is no single query behind them. • A verified figure is the right answer to the query that was run. If the question was understood differently than you meant it, the number can be correct and beside the point — which is why the answer cites its records and shows its queries. • Prose is not verified. Only figures are. • Verification is recorded for the answer you are looking at. Re-opening a past conversation from History shows the answer without a badge, rather than implying a check that isn't on file. Tips • The receipt is the fastest way to learn what the assistant can do - the queries it shows are exactly the ones you could build as a view or report. • An unconfirmed figure is often a sign the data moved between the answer being written and the check running, which is itself worth knowing. • Answers only ever read tables you can see, and hidden fields never reach the AI - so a receipt can never show you data you weren't already entitled to. Good to know • A verified number answers the query that ran, not necessarily the question you had in mind. Read the query in the receipt if the figure surprises you. • Derived figures (percentages, differences, rates) are not re-run and never carry a tick of their own. • If no model is configured, answers come from the built-in analyzer instead and are labelled as such; verification applies to AI answers.