ABOUT / VERIFICATION

How we check the numbers

Every number on EIGENROOM comes from code, and most published numbers are computed a second time by separate code in the test suite. This page says what those checks cover, when they run and what they do not cover yet.

Published answers are recomputed

The practice problems of all 51 lectures are checked by unit tests that recompute each numeric answer from the problem's data with their own routines: polynomial roots, Routh columns, partial-fraction residues, Runge–Kutta simulation, numerical integration, matrix exponentials, frequency sweeps, Nyquist encirclement counts and Riccati solutions. Each test then requires the page to state the value as rounded at the precision it shows.

The values in the lecture reference tables are recomputed the same way, independently of the code that prints them: Routh arrays, steady-state errors by system type, second-order specifications, Bode factors, lead phase, Ziegler–Nichols rules, root-locus rules, Laplace and z-transform pairs, Ackermann gains and LQR weights.

The worked examples of the eight calculation lectures (the inverse Laplace transform, second-order specifications, root-locus construction, Bode construction, the Nyquist criterion, PID tuning, solving the state equation and the z-transform) and the worked designs in the signal-flow-graph, lead–lag design and state-space design lectures are recomputed step by step, together with the lab numbers those lectures quote. In the eight calculation lectures a test also requires every decimal number in the lecture, its short answer and its FAQ, in English and in Korean, to be either a value the test recomputed or an input the text sets, such as a problem's data.

The same check now covers every other lecture as well: the body text, sections, formula notes, worked examples, short answers and FAQs of all 51 lectures are recomputed, more than 600 further values in all, using the labs' own engines wherever a lecture quotes a lab result, and each value must appear in English and in Korean at the precision shown.

Formulas, translations and links

  • Formulas. Every formula in the lectures (equations, symbol tables, worked steps, practice problems, short answers and reference tables) is rendered with KaTeX in strict mode during the tests. A formula that fails to parse, or that does not produce MathML for screen readers, fails the run.
  • Both languages. Every lecture section, practice problem, FAQ and journal note must have English and Korean text, and a lecture page that shows the same paragraph twice fails the tests.
  • Links. Links from lectures and journal notes into the labs must name an experiment that exists.

Simulations

  • Deterministic. Tests rerun the simulations and require identical samples, including runs with seeded sensor noise. The lab engines compute sine, cosine, exponential and arctangent with their own JavaScript functions instead of the browser's, so a shared run gives the same bits on every phone, computer and browser; a test fails if an engine calls a function whose last bit depends on the device.
  • Step size. Rerunning with a smaller integration step must agree with the published step within a small tolerance.
  • Exact references. The 6 Python reference scripts in the repository compare the engines with independent solutions: matrix exponentials for the DC motor, mass–spring–damper, state-feedback, RLC and system-type models; exact held-input solutions for the thermal controllers; SciPy integration for the nonlinear water tank and the two-degree-of-freedom bench; and polynomial roots for the Routh bench.
  • Fixed results. When the DC motor engine gained optional dead time and a right-half-plane zero in version 1.5.0, a test pinned every guided experiment and several older settings to the exact output of version 1.4.0, so links made before still replay bit for bit.
  • Stated definitions. Rise time, overshoot and settling time are defined on screen. A test restates each definition from that wording and compares it with what the engines report.
  • Field notes. The theory note behind each motor and cart-pole experiment predicts its run by hand. A test redoes each hand calculation, reruns the engine at the experiment's settings and at the two variations the note describes, and requires every resulting number in both languages; any other decimal in a note must be one of its stated inputs.

Calculators

The transfer-function calculators are tested against cases with known answers: the overshoot and peak time of a standard second-order system, the gain and phase margins of K/[s(s+1)(s+2)] at several gains, Z = N + P for loops that are unstable in open loop, root-locus axis crossings and breakaway points, Routh arrays with a zero first-column entry or a row of zeros, and Routh sign counts against numerical roots for random polynomials. The discretization calculator is checked against closed-form zero-order-hold results, a separate Taylor-series matrix exponential, the step response at the sampling instants, Tustin's exact match at the prewarp frequency and MATLAB's matched pole–zero result; the state-space calculator against the worked examples of the state-feedback and observer lectures, direct evaluation of C(sI − A)⁻¹B + D, and the eigenvalues of A − BK and A − LC. The numbers quoted in the calculator pages' explanations come from the calculators themselves.

When the checks run

  • On every push. GitHub Actions runs the type checker, every unit test, lint, the blog tests and a production build for each push to the main branch.
  • By hand. The Python reference scripts need NumPy and SciPy, so they are run by hand rather than on every push. So are the browser scripts (Playwright with axe), which open a local build and check the lectures in both languages, each lab's replay, sharing, import and export, phone layouts and accessibility.
  • After large merges. After the merges of 15 November 2026, a crawl of the built site checked the Korean pages for untranslated text; what it found is listed in the changelog.

Not covered yet

  • Model fidelity. A passing test shows that the code computes its stated model correctly, not that the model matches a particular machine. Each model's assumptions are on the method page.

Model and numerical method →

Found a wrong number?

Report it in the public feedback repository on GitHub. Fixes to published content are dated under Correction in the changelog.