What the transform actually does
Each pixel is converted from sRGB to linear light, multiplied by a 3x3 matrix, and converted back. The matrices are the ones published by Machado, Oliveira and Fernandes in 2009, derived from a shifted-cone-response model, and they are the same set most design and research tooling uses. Greyscale is not one of theirs — it is relative luminance, 0.2126R + 0.7152G + 0.0722B, with hue thrown away entirely.
The moderate and mild settings are a linear interpolation between the identity matrix and the full matrix. That is an approximation, and it is not identical to the severity-specific matrices in the paper. The full-severity output is the one to trust; the intermediate steps are a rough indication of the direction things move as a deficiency gets milder.
A simulator is a design check, not a diagnostic
Colour vision deficiency is not one condition. Anomalous trichromacy — a cone whose peak sensitivity has shifted — is far more common than dichromacy, where a cone type is absent, and the degree of shift varies continuously between individuals. Two people who both say "I am red-green colourblind" can have noticeably different experience of the same chart. On top of that, your monitor's gamut, its calibration and the ambient light all move the result.
So what this is good for is finding pairs of colours that are risky, and proving it to someone who is not convinced. What it cannot do is tell you what a particular person sees, or substitute for asking one.
The rule that matters more than any simulation
Do not encode meaning in hue alone. If red means failing and green means passing, add a word, an icon or a shape and the colour becomes reinforcement rather than the message. Chart series get direct labels or distinct line styles. Required fields get an asterisk as well as a red border. Map layers get hatching or texture. Status rows get "Overdue" written on them.
The payoff is wider than accessibility compliance. Anything that survives without colour also survives a monochrome printer, a projector with a blown colour channel, a screenshot dropped into a greyscale document, and a phone held at an angle in sunlight. It is a robustness rule that happens to also be an accessibility rule.
Palettes chosen with this in mind
Some sequences are designed to keep perceptual order under all three deficiencies — viridis, cividis and magma are the well-known ones from the scientific plotting world, and they hold their ordering in greyscale as well. For categorical series, blue against orange separates cleanly for protan and deutan viewers where red against green does not. If you have a free choice of palette, taking one of these off the shelf removes the problem before it starts.
How to work with it
Export the artwork at the size people will actually read it, drop it in, and look at the deutan panel first — it covers the largest group. Then look at greyscale. If both survive, the design is not leaning on hue. If either fails, the fix is almost never a different shade of the same two colours; it is adding a second channel of information, or moving one colour to a different lightness. Come back and re-run it after the change rather than assuming.
Questions people ask
Is this what a colourblind person sees?
No, and nothing can claim that. It is a model-based approximation of one type at one severity, rendered on your display. The most common form of deficiency is anomalous trichromacy, where a cone is shifted rather than absent, and the amount of shift differs from person to person. Use the output to identify colour pairs that are likely to fail, then, if the work matters, ask someone with the deficiency to look at it.
Why is my result blurry?
The image is scaled down to the long edge you set before the per-pixel transform runs, because the transform touches every pixel and a 6000 px photo is a lot of arithmetic in a browser. Raise the long edge to 2000 for a sharper result. Above that, memory pressure on phones starts producing blank canvases rather than slow ones, which is why it is capped.
Can I check just two colours?
The colour blindness preview tool is built for that — it takes a foreground and a background hex value and reports contrast ratio and colour distance in each state, with a suggested replacement when the pair fails. Use this simulator when you have a whole layout, chart or screenshot to judge.
Does the greyscale panel represent a real condition?
Complete achromatopsia exists but is extremely rare, on the order of one in thirty thousand. The panel is here as a design test rather than a simulation of a population: if information survives with hue removed, it is being carried by lightness, and lightness is the channel that every viewer and every reproduction method preserves.