Calculator
Verification debt calculator
Last updated:
Price it with your own numbers
Estimated cost of verification debt
Example – illustrative arithmetic, not a benchmark€67,000
per year · €5,580 per month
Modelled on about 120 AI-assisted merges a month.
At these inputs the model puts verification debt at €67,000 a year, or €5,580 a month: 74.4 hours of engineering time, 60 of them spent working out what a change was meant to do before it can be judged.
Hours per month
AI-assisted changes per month
- Review reconstruction
- Rework after a defect
- Break-even band, 30 to 40 changes a month
- Your current inputs
Modelled monthly engineering time against AI-assisted change volume, swept from zero to twice the current figure. The lower band is review reconstruction, paid on every AI-assisted change; the upper band is rework, paid only on the ones that turn out broken. Illustrative arithmetic with replaceable inputs, not a benchmark. Use the left and right arrow keys to move along the curve.
| Cost line | Per month | Hours a month |
|---|---|---|
| Review reconstruction | €4,500 | 60 h |
| Rework on churned code | €1,080 | 14.4 h |
| Incident allowance | none assumed | 0 h |
What the model assumed for you· Assumption set as of 2026-08-15
0.5 hours of review reconstruction per AI-assisted change · 2 % of AI-assisted changes reworked for a defect within 14 days, an illustrative rate to replace with your own · 6 hours to rework one churned change
Find your shape
What verification debt costs, by team shape
Six shapes, priced by the same arithmetic the calculator above runs. Find the row that looks like your engineering organisation, then move the sliders to make it yours. What decides the figure is volume, not headcount, which is why a ten-person agency shipping mostly AI-assisted changes carries more debt than a thirty-five-person department that barely uses it.
| Team shape | AI-assisted changes | Engineering time | Modelled cost | Does verification pay? |
|---|---|---|---|---|
| Solo founder or a pair2 engineers who merge · 60% AI-assistedEarly product, one repository | 24 per month | 14.9 h per month | 13,000 €per year1,116 € per month | Below the band — probably not |
| Seed-stage startup, one squad6 engineers who merge · 60% AI-assistedFirst hires, shipping daily | 72 per month | 44.6 h per month | 40,000 €per year3,348 € per month | Above the band — yes |
| Agency or development shop10 engineers who merge · 80% AI-assistedClient projects, high AI adoption | 160 per month | 99.2 h per month | 89,000 €per year7,440 € per month | Above the band — yes |
| Scale-up, three squads18 engineers who merge · 60% AI-assistedSeries A to B, several services | 216 per month | 133.9 h per month | 121,000 €per year10,044 € per month | Above the band — yes |
| Mid-size company IT35 engineers who merge · 40% AI-assistedInternal systems, cautious adoption | 280 per month | 173.6 h per month | 156,000 €per year13,020 € per month | Above the band — yes |
| Enterprise platform group80 engineers who merge · 40% AI-assistedMany teams on a shared platform | 640 per month | 396.8 h per month | 357,000 €per year29,760 € per month | Above the band — yes |
Why the bill exists
The expensive part is not the code that breaks
A reviewer looking at an AI-assisted change has to answer a question that used to answer itself: what was this supposed to do? When a colleague wrote it, the intent arrived with the change, in a ticket, a conversation or the shape of the commit. When a model wrote it, the reviewer reconstructs the intent from the diff before they can judge whether the diff is right.
That reconstruction is paid on every change, not only the ones that turn out to be wrong. It produces no ticket and no defect record, so it never lands in a budget line, and it is the reason a team can feel slower while shipping more. At the default inputs above it is four fifths of the total.
The rest is ordinary rework, and it is worth being precise about what can and cannot be claimed there. Industry churn studies measure lines of code revised after merge, which is a useful warning signal and a poor defect rate. This model does not convert one into the other; it asks you for a rate your own tickets support. The full cost model sets out both the formula and what it refuses to do, and the measurement guide shows how to get the inputs from your own repository rather than from an average.
Being built
What a Team version changes
Coming soonNot built yet. Listed so you can judge the direction, not the delivery date.
Reality Graph is a single-operator tool today: no accounts, no seats, no shared state. Four things change when that stops being true. None of them is running yet, all of them are dated, and the price is published so you can plan against it.
One checklist every change is held to, set once for the team
Right now each engineer decides for themselves what a change has to prove before it counts as done, so two people can both say verified and mean different things. One shared checklist makes their answers comparable, which is what lets you trust a review you did not run yourself.
Today: Write the checklist down and agree it in the team. Nothing enforces it yet, but the argument about what counts is worth having early.
A review history a second person can open
The engineer who ran it stops being the only person who can answer what happened. That is the difference between evidence and a story about evidence.
Today: The record is a plain file on the machine that produced it. You can read it, and you can send it.
Roles and approvals across a team
An approval stops meaning I approved my own work. Who may sign off on what becomes a property of the organisation rather than of the terminal it was typed into.
Today: Approvals are bound to the run they were given for, so at least none of them becomes a standing permission.
Central configuration for a whole organisation
One place to set what every project must check, instead of a convention that survives exactly as long as the person who wrote it stays.
Today: Configuration is per project, which is workable at three repositories and painful at thirty.
39 euro per user, per month
The intended price. Reality Graph has no accounts or seats today, so nothing can be bought yet.
Expected
End of this year
not built · no accounts, no seats, no shared state
Free covers commercial use, so nothing stops a company running it on every machine today, and that is deliberate: a startup should be able to stabilise its own product without asking anyone. The four changes above are the only reason a Team version would be worth paying for, and none of them exists yet. Telling us which one you would use first is what decides the order they get built in.
Where to go from here
Two directions, depending on which question you are still holding: what the problem is, or what the tool does about it.
- ArticleVerification debt, definedThe term, where it came from, and the survey data behind it. Start here if the calculator's premise is the part you want to check.
- CapabilityThe deterministic verdictThree outcomes, computed from evidence rather than from a model's opinion, and what a pass does not mean.
Questions people actually ask
- Where do the numbers in this calculator come from?
- From the worked example published on the cost model page, not from a study about your industry. Two of the inputs are yours to supply, the rework rate and the reconstruction time are illustrative values the model shows openly, and the arithmetic is the same one the cost article prints. Nothing here is a measured customer outcome, and the page says so beside the figure rather than in a footnote.
- Why is the largest line one nobody budgets for?
- Because review reconstruction is paid on every AI-assisted change, not only the ones that turn out to be broken. Working out what a change was meant to do before you can judge it does not produce a ticket, a defect record or a line item, so it never appears in a budget. At the default inputs it is four fifths of the total, and it is the reason the number surprises people.
- Does this calculate what Reality Graph would save us?
- No, and it will not until there is a before-and-after measurement to base it on. No such measurement exists, so any percentage here would be invented. What the page does instead is name which cost line a verification practice is designed to address, cite the one published figure on that side, and say where the practice stops paying at all.
- Our team is small. Does verification still pay?
- Below roughly 30 to 40 AI-assisted changes a month, probably not. Volume decides it rather than headcount: two people running agents all day are past that threshold and twenty people barely using AI are not. The calculator shows the band and says plainly when your inputs fall under it, because an argument that cannot lose is not an argument.
- Can I buy the Team version today?
- No. Reality Graph has no accounts, no seats and no shared state at all, so there is nothing to buy yet. The Team price is published so you can plan against it, and the waiting list is a conversation about which part you would use first, which is what decides the order things get built in.
- How much does unverified AI-generated code cost a 20-person engineering team per year?
- At twenty engineers merging around twenty pull requests each, with three fifths of them AI-assisted, the model puts it near 134,000 euro a year at a loaded rate of 75 euro an hour. Four fifths of that is review reconstruction rather than fixing defects. Treat the figure as arithmetic you can check rather than a measurement of your team: the table of team shapes above shows the same model at six sizes, and the sliders replace every input with yours.
- What happens to the cost when our AI-assisted change volume doubles?
- It doubles. Both cost lines in this model are linear in volume, which the chart draws as straight edges rather than a curve, and that is the useful part: nothing accelerates at a threshold, so the debt tracks adoption rather than ambushing you at some size. It also means the reverse holds, and halving the share of merges that go out unverified halves the bill.
- How do I turn this estimate into a number my CFO will accept?
- Replace the two inputs that are yours before you quote it: the loaded hourly cost your finance team already uses, and a rework rate taken from your own tickets rather than from an industry study. Then quote the hours rather than the euro, because hours are what you reallocate and they survive an argument about rates. Bring the assumption line with you; a figure whose method is visible is much harder to wave away than a round number.
- Do we have to send you anything to use the calculator?
- No. The arithmetic runs in your browser and the numbers you type never leave it. The site records that the calculator was used and a rounded estimate, both without any identifier, and never the team size or the hourly rate you entered.
Sources
Tell us which part you would use first
The Team version is not built. Which of the four changes above would matter most to your team is what decides the order they get built in, and saying so takes a minute.