Editorial policy

What decides, and what only explains?

The engine decides; the prose explains. Every dollar figure on these pages comes from deterministic code with zero dependencies, no network access and no AI anywhere in the calculation path. Money is held in integer cents and rates in basis points, so a phase-out computed at the same MAGI returns the same cents every time. The writing can describe a result or warn about it; it cannot move it.

Where do the rules come from?

From the statute as enacted and the forms that compute it. The four deductions are read from P.L. 119-21 itself — sections 70103, 70201, 70202 and 70203, which add IRC §§ 151(d)(5)(C), 224, 225 and 163(h)(4). The computation of record is Schedule 1-A (Form 1040) and its line-by-line instructions, because that is what a return is actually filled in against.

Four further primary documents carry specific pieces: the 2026 inflation-adjustment revenue procedure for the bracket table and standard deduction; the final tips regulation for the occupation list, the phase-out ordering and the confirmation that the $25,000 cap is per return rather than per spouse; the IRS notice and fact sheet that are the operative guidance on qualified overtime; and the W-2 instructions that establish where an employer reports these amounts. Nothing enters the engine from secondary coverage. Every citation and its last-verified date is on the Sources page.

What closes a verification gap?

A primary source, read directly. A secondary source may cross-check a figure and may never close a gap. Where a source could not be reached, the row stays open and the engine keeps whatever it already had — no value is filled in from recall, and no row is quietly improved with a plausible substitute.

One method rule comes from a near miss rather than a principle. Figures are read from statute text or from PDF text extracted locally, never from a summarising layer. During this same verification programme, a summarising fetch tool misnumbered a revenue procedure and returned a completely fabricated table for another engine on this site. The encoded values were already correct; acting on that summary would have replaced a right answer with garbage.

Blocked sources are disclosed rather than papered over. Several official hosts refuse automated requests, so where a citation points at a mirror or at the enrolled bill instead of the code section, it is because that is the document that was actually opened. No source relevant to these four deductions was lost to a block, with one exception worth naming: no Treasury regulation implementing the overtime section could be located, and because the Federal Register's own search could not be browsed, its absence is not established.

What review gate has not been met?

Credentialed review. Tax logic must be reviewed by a credentialed reviewer — an enrolled agent or a CPA — before launch and after every rule change, and their name and credentials must appear on the pages they signed off.

How are errors corrected, and how fast?

Within 48 hours of confirming one, in a fixed sequence: correct the rule file, add a regression test that fails on the old behaviour, re-run the suite, then publish a dated changelog entry naming the source the change was made against and which results moved.

Two corrections already on that record show the standard. The tips and overtime phase-outs were rounding the $1,000 step up where Schedule 1-A lines 11 and 19 require it rounded down, which understated both deductions by up to $100 each. And the senior deduction applied its 6% reduction once to a doubled amount, where the form applies it per person and enters the result twice — a joint return with two spouses over 65 at $200,000 of MAGI was returning $9,000 against the form's $6,000. Both were fixed and both are now pinned by tests.

Report an error to info@bracketsight.com, with the inputs you used if you can. We will credit the report unless you ask otherwise.

How are rule changes versioned?

One dated file per rule set per tax year, each carrying its own version string, an effectiveFrom and effectiveTo date, a verified flag, and citations with the date each was last checked. Six files back this section: one for each of the four deductions, one for the bracket table and standard deduction, and one for the occupation list. The bundle version shown on these pages is derived from the files rather than typed by hand, and the date beside it is the oldest last-verified date across the whole bundle — the honest figure, not the flattering one.

A new tax year gets new files, never an edit in place, and these files must not be cloned past tax year 2028: the tips deduction allows nothing for a taxable year beginning after 31 Dec 2028, and the senior and car-loan deductions carry the same sunset. A file that outlived its statute is worse than a missing one.

Who pays, and can they reach a number?

This section is planned to be advertising- and affiliate-supported. No advertising runs today and no ad network script is loaded on any page. Advertisers and affiliates cannot influence a computed result, and the guarantee is structural rather than a promise: the engine has no dependencies and makes no network calls, so there is nowhere inside a calculation for a partner's code or data to sit. Every formula is published in full on the Methodology page, and the arithmetic runs identically whether or not any partner exists. No ad slot sits inside the tool, where it could be mistaken for part of a result.

What AI does — and doesn't do

No AI computes, estimates, or adjusts any number on this site. Every figure comes from deterministic, unit-tested arithmetic over cited rules. Any future AI feature, such as a plain-language explanation, will be validated number-for-number against engine output and will fail closed to a deterministic template.