Every answer ships with a receipt of the exact SQL.
Natural-language-to-SQL is only trustworthy if you can see the SQL. at1 askreceipt answers questions in plain English and attaches a signed, re-verifiable receipt of the exact query it ran — tables, columns, row count, result hash. A wrong-column answer stops being a plausible number and becomes a visibly-wrong query an auditor can inspect and re-run.
- receipt per answer
- the exact SQL, signed and re-runnable, ships with every result
- tables & columns
- the receipt names what was read — a wrong column is now visible
- result hash
- re-run the SQL and the hash must match, or the answer is stale
- re-verifiable
- an auditor checks the query, not your word for the number
Ask in plain English
"What was total revenue for branch 7 last month?" is turned into SQL and run against your data — the same as any NL→SQL tool, up to here.
Get a receipt
The answer arrives with a signed receipt: the exact SQL, the tables and columns it touched, the row count, and a hash of the result set.
A wrong answer is visible
If the model read the wrong column, the receipt shows it read the wrong column — the mistake is in the query you can read, not hidden inside a plausible-looking number.
The workflow
# ask in plain English -> answer + a signed receipt of the exact SQL at1 askreceipt ask data.at1 "total revenue for branch 7 last month" -o answer.json # re-verify the receipt: re-run the SQL, check the result hash still matches at1 askreceipt verify answer.json # show the receipt: the SQL, tables/columns touched, row count, result hash at1 askreceipt show answer.json
Honest scope
The receipt makes the query auditable and reproducible — it does not make the SQL correct. The model can still misread your intent; what changes is that the mistake is now in a query you can read and challenge, instead of buried in a number. The re-verification proves the same SQL over the same data gives the same result. The right check is your acceptance test on real data — feed it your trickiest questions and read the receipts. Validated on synthetic schemas; partner acceptance on your data is pending.