Today's red-team run aimed at two finance tools that read a CSV someone pasted: Rev Rec (a deferred revenue schedule from invoice lines) and MRR Pulse (MRR, churn and cohorts from a subscription export). We looked for inputs a stranger might paste that give a wrong number and no warning. We found three.
1. 30 February becomes 2 March
Date.parse("2026-02-30T00:00:00Z") does not fail in V8. It returns 2 March. Both importers passed dates through Date.parse, so a typo in a period start moved the revenue into the next month, and nothing on the page said so. Most of the repository's date parsers already check that a day round-trips. These two didn't. They do now: a day that does not exist marks the row as skipped ("period start is not a date"), the same as any other unreadable date.
2. € 1.000 read as one euro
Rev Rec already read 1.234,50 as 1234.50, because a comma followed by one or two final digits is a decimal comma. A whole-euro amount with a dot as the thousands separator, € 1.000, has no comma. The importer read it as 1.000, which is one euro, so that row came out a thousand times too small. Nothing in the text can tell "one thousand" apart from "one, to three places". But a two-decimal currency is never written to three places, so the row is now skipped: "amount has more decimals than EUR allows (is . a thousands separator?)". Padding zeros such as 5000.00 yen still import. A non-zero extra digit, like 100.5 yen, is skipped.
3. A mistyped end date reads as "still a customer"
In MRR Pulse, an ended at value that did not parse (say soon, or 2026-02-30) came back as "no end date". The rebuilt history then counted that canceled subscription as alive every month since it started: no churn, and inflated MRR for each past month. The headline number was right, because it reads the status. The history was wrong. A canceled row with a blank end date did the same. All of these rows are now skipped with a reason, and so is a trial-end date that does not parse.
The tests that were missing
The run also seeded 20 small faults into the two tools' core maths. The test suites caught 15. Among the five that got through: a one-cent month left out of the journal, a rate shown as NaN instead of "n/a" when there was nothing to divide by, and a sort that stopped ranking the largest customer first. Each of the five now has a test that fails when the fault is put back, and the catch rate on those 20 faults went from 75% to 100%.
The tools are free and MIT licensed. They give a working schedule and working metrics for your own bookkeeping. They are not accounting, tax or audit advice.