The useful feature is a decision that can explain itself.

Calculation features

Exact cents, independent bases, explicit rounding, and a net-pay floor.

Gross-to-net keeps amounts in integer cents from intake through result. Taxable-basis construction applies a signed exemption mask separately for each authority. FICA evaluates exact integer ratios and names the rounding boundary. Federal withholding is designed to follow the selected signed worksheet rather than an approximate percentage. Deductions retain authority treatment. The published gross-to-net primitive floors net at zero and flags over-deduction instead of hiding the deficit.

Those mechanics are intentionally composable. An operator can inspect the basis before a rate is applied, compare a withholding result with the source worksheet, or isolate a deduction from the rest of the run. Small functions make it possible to identify where disagreement began. They also keep a missing statutory table from being masked by successful arithmetic elsewhere.

A worked feature trace

Follow one deduction through basis construction, withholding, composition, and reconciliation.

Take a deliberately synthetic earning statement expressed entirely in cents. The trace begins by validating every amount and reading the signed exemption mask. Instead of subtracting one combined pre-tax total, basis construction evaluates the deduction independently for federal income tax, Social Security, Medicare, and each configured jurisdiction. The trace returns the included and excluded components per authority, making a classification disagreement visible before any rate is introduced.

Federal withholding then selects the signed worksheet whose effective window contains the pay-period boundary and exposes the worksheet steps. FICA applies its exact rational rates to the relevant bases, separately identifies employee and employer sides, respects the Social Security wage limit supplied by the governing rules, and treats Additional Medicare as employee-only. Fractions are handled at the named rounding point; binary floating-point is not allowed to choose a cent accidentally.

Gross-to-net composes those primitives rather than maintaining a second private formula. That means the federal result inspected alone must equal the federal component in the composed statement. The computed net is floored at zero and an over-deduction remains flagged; the return does not recover a shortage from a later period. Each output retains enough intermediate evidence to locate which primitive produced it.

Finally, the reconciliation feature compares the resulting authority totals with synthetic quarterly and annual artifacts. A one-cent or one-box mismatch remains a mismatch. The trace does not introduce a suspense amount, auto-correction, or probabilistic explanation. This worked path illustrates why the feature inventory emphasizes composability, provenance, integer arithmetic, and refusal codes: those qualities let an auditor reproduce a decision without treating the application as the filing or disbursement system.

Test a feature claim with both an answer and a counterexample.

A fictional evaluator writes an acceptance sentence for independent bases: changing one authority's exemption treatment must not silently change another authority's basis. Beneath it, they write a counterexample with two invented authorities and one deduction that applies differently. The paper exercise does not select a legal classification. It shows what evidence the feature needs to expose. A single taxable total would fail the explanation even if it happened to equal one authority's expected value. This is why a useful feature description names an invariant, not merely a screen.

For exact-cent composition, the evaluator compares the federal component inspected separately with the federal component placed in the composed result. They are asking whether the same primitive contributes the same amount, rather than accepting two unrelated formulas with similar labels. A difference would need investigation before anyone approved an explanation. This website does not run that comparison for the visitor. The published engine reference identifies the component boundaries; the manual exercise describes the evidence a reviewer would ask to see in a scoped evaluation.

Capability, reachability and external completion need separate evidence.

Use a manual three-column ledger when reviewing the feature inventory. The first column records the named calculation or validation contract. The second records whether the published build ledger identifies a caller. The third records the action the feature does not perform. A pure validation can occupy the first column while the second remains built without a production caller. A structural bank preview can have a caller while the third still says raw file withheld and nothing transmitted. Neither condition should disappear into a general available badge.

An evaluator may prefer one yes-or-no feature list. That is understandable, but a single checkmark would combine several different promises. Wages can describe exact arithmetic without offering a payroll service, and it can name a built module without offering a public control for it. There is no AI assistant, live checkout or disbursement action in these marketing guides. The useful answer is a narrow capability and its stopping point, with source-backed limits retained. A broader workflow requires separate evidence; the inventory does not supply it by implication.

A final manual comparison can distinguish a feature defect from a changed example. Keep the original invented inputs beside the revised inputs, and name the one field that changed. If both the classification and amount changed, the comparison cannot isolate either cause. Do not describe the exercise as a saved history or an automated regression runner: it is a reviewer working with two written examples. Its value is an explicit question about the contract, with the unchanged assumptions visible. Reproducibility begins with that discipline, even before software is involved.