How the calculators are tested

The About page claims the arithmetic on this site is checked before it ships. This page is that claim in table form: every test suite in the build, what it re-derives, and how many separate assertions it makes. The numbers below are not typed by hand. They are written into this page by the build script from the output of the suites themselves, so a suite that shrinks, breaks, or disappears changes this table on the next release.

Last built from a full run of the suites: 8,451 assertions across 34 of them, 31 mirroring one calculator each and 3 covering the guides and the site’s own published claims.

What one assertion is

A mirror suite is a second implementation of a calculator, written in Python, that never imports the site’s JavaScript. It computes the same answer from the same inputs and compares. One assertion is one such comparison: this input, on this scale, must produce this output. A tool with 60 assertions has been asked 60 questions with known answers — ordinary cases, boundary scores, impossible targets, already-secured targets, malformed input, and the exact rounding the browser performs, which is not the rounding Python performs by default.

Where a calculator carries a published table — a letter ladder, a percentage band, a classification boundary — the mirror reads that table out of the JavaScript source and asserts it against its own copy, so a silent edit to a grading scale fails the build rather than shipping.

The suites at the foot of the table are not tied to one calculator. They re-derive the worked examples printed in the guides that have no calculator of their own, and they check the site’s own statements: that the counts on the homepage match the number of files, that the pages promising a source actually link one, and that the fetch log behind every outbound citation records a 200, from its latest check or, where the linked site’s robots.txt or bot defenses rule out a re-check, from the dated check that last confirmed it.

What this table does not prove

Two implementations that agree still rest on one reading of the source. A mirror catches arithmetic error, rounding drift, and a table edited in one place and not the other. It cannot tell you that a policy has been read correctly, that an institution has not changed its rules since the citation was checked, or that a calculator is answering the question you actually have. Those are what the sources on each page and the plain statements of limits are for, and they are why every tool says the same thing at the bottom: your syllabus wins.

The suites also run before a release, not in your browser. If you find a number here that disagrees with your own arithmetic, that is worth an email: a confirmed error becomes a failing assertion in the relevant suite first, and only then gets fixed, so the same mistake cannot return in a later change.