You don't pay for EthicGuard. You stop paying for rework.

Utrecht University analysed 1,023 user stories from 18 organisations and found 56% carried at least one defect; their earlier manual pass put it at 24%. We model 1 in 8 — well under both — because those studies count structural writing smells, most of which are cheap to fix, and we count only defects severe enough to cause rework. Ours · reasoned, not measured No study reports 12.5%, and we attribute it to nobody. The quantity our model actually needs has never been measured; a rate from real AC-defect labels will replace it and will beat every citation on this page.

Porter, Votta & Basili (University of Maryland, IEEE TSE, 1995) measured requirements inspection finding between 24% and 57% of faults. We sit near the bottom of that band: EthicGuard is an automated grader, not a trained inspection team. Duisburg-Essen ran an LLM over 1,266 annotated Mercedes-Benz requirements and got recall above 0.98 at precision near 0.55 — it sees nearly everything and over-reports badly, which is exactly why a human keeps the final say. The other two thirds are the figure in the right-hand column.

Only the difference is counted — writing the criteria costs something either way. A single-digit multiple, not the "100× after release" curve: that curve is challenged by Menzies et al. (NC State / SEI Carnegie Mellon / Fraunhofer, 2017), and the 1:10:100 figure most calculators quote traces to no primary source at all. Wagner (TU München, 2006) puts removal at 1.06 h at requirements review and 8.37 h at system test — a 7.31 h delta, which is where our 7 comes from. Two caveats we would rather state than have found: that paper is a literature survey, not a primary measurement, and it compares techniques rather than following one defect through time. Frattini et al. (2023) reviewed 57 papers and found this cost is reported in 15.8% of them and “never determined empirically”. Nobody has measured it — us included. Hours per developer-year are taken as 1,800, after leave and non-project time.

List price on the Atlassian Marketplace. The ladder is marginal, not flat: a 250-user site pays 100 × $9.05 plus 150 × $7.65, not 250 × $7.65. This page walks that ladder rather than multiplying by a headline rate, so it can never quote a price the listing does not charge — and sites of 10 users or fewer pay nothing at all.

Counted as a cost, because it is one. Against the licence alone this page reported a 1,844% return — arithmetically true, and the same meaningless number every vendor calculator produces by dividing savings by list price. Adopting a tool that reviews 960 stories a year is not free because the software is cheap. Five minutes is generous to the cost side: most verdicts are a clean pass. No study publishes this — it is ours to measure from real usage, and it should be replaced the moment we can.