Questions answered from the metric catalogue over your own data · every number in an answer is a cell in the result · the system says what it cannot answer
Ask
Green answered · amber partial (coverage gap named) · red declined (no data)
Spend and matters
Firms and rates
Exposure and outcomes
Compliance
Pick a question. The answer shows how it was interpreted, the rows it was computed from, the narrative built only from those rows, and what would make a partial answer complete.
How an answer is produced
Why this does not hallucinate: the language model never writes the number, it only chooses the metric and explains the rows
1 · Interpret
The model maps the question to named metrics, grains, filters and a time window from the catalogue. Unknown terms are asked back, not guessed ("by 'exposure' do you mean recorded claimed amount, or reserve?").
2 · Plan
The plan is parameters against the catalogue (metric, group-by, filters, window), validated against the schema and the user's role scope. The model does not write SQL; the query engine does.
3 · Compute
The warehouse runs the query and returns rows with n, coverage and freshness. Cohort comparisons below the minimum n return an abstain instead of a number.
4 · Narrate
The model writes the narrative from the rows. A checker verifies every number and entity in the text appears in the rows; anything that does not is removed and the answer is flagged.
5 · Ground and limit
The answer carries the plan, the rows, coverage, sources and limits, and links to the underlying matters and lines. Partial answers state the gap and the fix; declined answers state why.
