The second reader you wish you had before you hit send.
Nothing to paste? — a weekly revenue review with a headcount recommendation attached and a few of the usual problems buried in it. Or , rendered from a saved reply with no run and no charge.
Tables, query, method notes, the conclusion you plan to lead with. The number scan runs in your browser as you type — free, no sign-in, nothing sent anywhere.
The question and the audience change the audit. An analysis that is fine as a team update can be unfit for a decision meeting, and the verdict says which.
Seven named failure modes, each marked clear, suspect or found with the evidence attached. Anything found rules out "Ready to share" — no exceptions, no small print.
The checklist puts the blocking items first. When you have revised, re-audit — the second pass is compared against the first, so you can see what closed, what is still open and whether the revision broke anything new.
Seven named failure modes, drawn from @anthropics/data-validation. Each is checked against what you paste and marked clear, suspect or found — with the evidence quoted, so you can disagree.
The population a rate is measured against changes between the periods being compared, so the rate moves even though behaviour did not. A conversion rate over “active users” falls if the definition of active widened and rises if it narrowed, whoever converted. The fix is to state the denominator explicitly for every period and confirm it was computed the same way each time.
It gives every group equal weight regardless of size, so ten customers count as much as ten thousand, and the result describes no real population. Compute the weighted mean from the underlying totals instead — sum the numerators, sum the denominators, divide once.
A query joins on a key that is not unique on at least one side, rows are duplicated, and every sum, count and average downstream is inflated. The signs: a row count larger than the source table, revenue exceeding what the business actually made, counts that grow when a purely descriptive table is joined in. Check the join key is unique on the joined side before trusting any aggregate.
Drawing conclusions from only the cases that made it into the data, when the ones that
dropped out carry the information you needed. Retention computed over customers who are
still active today will look excellent by construction. The usual form is a status filter
— status = 'active' — applied at query time to a
historical window.
When the methodology supports the conclusion, no pitfall was found, no calculation check failed, and the remaining caveats are ordinary scope notes rather than corrections. If any pitfall was found or any check failed, it is not ready — there is no minor qualifier that gets around that. The middle state, share with noted caveats, means it is sound enough to circulate but only if the caveats travel with the numbers.
Any limit on what the numbers mean: the population and window they cover, metric definitions that would change the answer if read differently, known gaps or exclusions, and comparisons whose baselines are not strictly like for like. A caveat is not a disclaimer protecting the analyst — it is what a reader needs in order not to over-read the result, which is why it has to move with the numbers.
The arithmetic that is visible in what you paste, and it says so. Two layers: a free browser-side number scan flags percentage groups that do not sum to 100, total rows that disagree with their own column, percentages above 100, negative counts, duplicate rows and empty cells; and the audit runs calculation spot-checks marked PASS or FAIL with the working shown. A number whose derivation is not on the page becomes an open question, not a silent pass.