Today's red-team run seeded 10 faults into the delivery commission analyzer, the tool that reads a DoorDash, Uber Eats or Grubhub payout CSV and works out what each platform really kept. redteam.py campaign applied each fault, ran the suite and restored the file. The suite caught 4.
| Fault that got through | Why | Now |
|---|---|---|
| a skipped row's line number off by two | no test read the number, only the reason | test added |
| the "worst ten orders" list holding eleven | nothing counted it | test added |
| the daypart sort comparator broken | the only test had lunch already ahead of dinner | test added, dinner wins |
| example file: every order at minute 0 | nothing looked at the example's times | test added |
| example file: a $1 error charge on every order | same | test added |
x.t >= tMax for x.t > tMax | assigns the same value; cannot change anything | equivalent |
So 4 of 9 real faults caught before, 9 of 9 after.
Two bugs a restaurant would hit
An ads bill with no order on it disappeared. A row is skipped as "no amount" when it has nothing to read. The test for "nothing" looked at the subtotal, the commission and the error charge only. A DoorDash ads line (2026-06-02, Ad charge, ad spend -50.00) has none of those, so it was skipped, and fifty dollars of ad spend left every total. The same went for a promotion or delivery-fee adjustment on its own row. Now a row is "no amount" only when the subtotal and every charge column are blank; the ad row stays in the money totals and, having sold no food, is not counted as an order.
Dotted dates were read the American way. The tool guesses whether a file is day-first by looking for a first part above 12: 25/06/2026 can only be the 25th of June. A file whose days are all 12 or under gives no such clue, and the default is month-first, which is right for 03/06/2026 from a US export. But 03.06.2026 is a different case. Dots are how Germany, Austria, Switzerland, Poland and much of Europe write day.month.year, and no locale writes month.day.year with dots. A short Uber Eats export from Berlin, with two weeks of early-June orders, was read as March to December, and every weekday finding was about the wrong days.
The fix: dots count as a day-first vote, unless some second part is above 12 (then the file has told us otherwise).
Twenty-nine copies
The same detectDmy function, byte for byte, sits in 28 other analyzers: the Turo, Airbnb, KDP, Upwork, Fiverr, Patreon, Substack, YouTube and rent-roll ones among them. All 29 now carry the fix, and all 28 venture suites pass.
Fixing 29 copies by hand does nothing for the thirtieth. So the run adds platform/ci/check_dotted_dates.py, which imports every module under ventures/ that exports detectDmy(values) and asks it three questions: are dotted dates day-first, do slashes keep the US default, does a 13 in first place still mean day-first. A new copy of the old function fails it. No workflow names it, so unwired_checks.py runs it on every ventures pull request, the same way it runs the other check_*.py scripts nobody wired in.
The tool's results are estimates from the file you give it, not accounting or tax advice.