How we verify every figure

Most calculator sites ask you to trust them. This one publishes the mechanism instead, because a claim you can check is worth more than a credential you cannot.

Right now the site carries 10 calculators backed by 69 hand-verified test cases. Every one of those cases runs on every build. A single failure and the renderer writes nothing at all — the previous version of the site stays up rather than a wrong number replacing it.

What runs before anything publishes

The build has a gate between the source and the rendered page. It runs six passes, and the renderer refuses to start if any of them reports a problem.

1. Purity. Each calculator’s logic is scanned to confirm it is a pure function that behaves identically in a browser and on a server: no clock, no randomness, no network, no file access, no hidden state. A calculator that reads the current date could give two readers different answers to the same question.
2. Golden cases. Every calculator ships with test cases whose expected figures were worked out by hand, line by line, against the relevant IRS form — independently of the code. The code has to reproduce them to the cent. Each case records who verified it, on what date, and by what method.
3. Property testing. Two thousand randomised input combinations per calculator, drawn from the declared range of every field, checked against invariants the calculator declares about itself — that the figures on the page reconcile with one another, that the answer moves in the right direction when an input rises, that no combination produces a nonsensical value. The random draws are seeded, so any failure can be reproduced exactly.
4. Citation coverage. Every tax figure a calculator reads is traced at build time. If a calculator reaches a rate or threshold that has no citation attached, the build fails. An uncited number cannot reach a page.
5. Content checks. Each page must carry a minimum of genuine explanation, a stated method, an explicit list of what it does not model, and a verification date within the last year.
6. Duplication. Every page is compared against every other on five-word sequence overlap. Pages that read too similarly fail the build, which forces each one to be written rather than templated.

The most similar pair of pages currently on the site shares 2.5% of their five-word sequences. The build fails above 35%.

Why property testing, not just examples

Hand-checked cases prove the calculator is right about the situations someone thought to check. They cannot prove it is right about the situations nobody thought of, and those are where calculators actually break.

The randomised pass is what covers the gap, and it earns its place. Both calculators on this site shipped with a defect that the hand-written cases passed straight over and the random draws caught immediately — in one case a figure that displayed as zero while simultaneously advising the reader to sell a share to cover it, in the other a calculator that reported a limit of zero to someone whose tax situation had nothing to do with the option they were asking about. Neither was exotic. Neither would have been noticed by a reader, who would simply have acted on a wrong number.

Where a number is not good enough, we do not publish it

The tax rules are stored with a status attached, and some figures are explicitly marked unverified: we have a plausible value, but we have not confirmed it against a primary source to the standard above.

The build refuses to let any page read an unverified figure. That has real consequences — the calculators currently do not offer *married filing separately* as an option, because the 2026 alternative minimum tax phaseout threshold for that status is genuinely ambiguous between two readings of the statute and we will not guess on a money page. Tax year 2025 is absent for the same reason. Both will appear when they have been checked, and not before.

A missing option is an honest answer. A confidently wrong one is not.

What we cite

Sources are the Internal Revenue Code, Treasury regulations, IRS forms and publications, and the annual Revenue Procedure that sets the inflation-adjusted figures. We link to the statute at Cornell’s Legal Information Institute and to the IRS directly. We do not cite other commentary sites, including well-known ones, because a citation chain that ends at somebody else’s blog is not a citation.

What this is not

Verification is not advice. A calculator that is arithmetically perfect can still be the wrong tool for your situation, and none of these model your full return. Equity compensation interacts with state tax, with the rest of your income, with your filing status and with timing decisions in ways a single-purpose calculator cannot see.

Everything here is educational information. Before acting on any figure of consequence, take it to a qualified tax adviser — ideally one who has seen equity compensation before.

Check it yourself

The machine’s own output is published at /test-report.json: every calculator, every test case counted, every invariant named, every tax-rule path each calculator reads, and the duplication score. It is regenerated on every build.

If you find an error, tell us. A correction is worth more to this site than the page it corrects.