Jim Paris has made a real TrueType font. You type [hello] into a text field, apply the font, and the brackets become a scannable QR code. There is no image-generation step and nothing computing the code off to one side. The QR code is drawn during shaping — the same stage of the text pipeline that fuses f and i into fi, the same machinery that joins Arabic letters to each other. And because the result is still text, you can select the QR code, copy it, and paste it into a plain text file. What you carry away is the string.

There are three fonts: QR Font 1-L (21×21 modules, up to 17 characters), 2-L (25×25, 32), 3-L (29×29, 53). None of them was drawn by hand. A Python script emits the glyph outlines and the GSUB feature logic, and a compiler turns that into a .ttf. The ordinary readable text outside the brackets is Liberation Sans, scaled into the QR em square — "Liberation" is a reserved name under the licence, so he had to rename his families.

It is worth stopping to see where this font's weight sits. In a Version 1-L symbol, 152 bits are data and 56 bits are Reed–Solomon parity. The 56 carry nothing. They are the remainder of a polynomial division, and their only job is to be there so that when a corner is torn off, or a cup is set down on the sticker, the rest can reconstruct what is gone.

Almost all the machinery in Paris's font exists to produce those 56 bits. Scanning characters into bytes, filling the length field, expanding to bits — cheap. The parity, by his own design notes: 152 large contextual lookups, roughly four thousand toggle actions, a few thousand intermediate glyphs. The README even ships a build target that skips it — make fast-placeholder, "a faster layout-only build with placeholder zero parity." That font still draws a lattice of black squares. It simply doesn't scan.

Which is to say: what makes a QR code a code is the part of it that says nothing. And this typeface, measured by machinery, is almost entirely that part.

The method itself is elegant. He does not implement GF(256) polynomial division inside the font; he linearises the problem outside first. For a fixed version and error-correction level, the 56 parity bits are a linear transformation of the 152 data bits over GF(2). Feed a reference encoder 152 test vectors, each with a single bit set, record the 56 parity bits that come back, and that is column i of the matrix. Then every data bit compiles down to one contextual rule: when D37 is 1, toggle P2, P5, P8, P11, and so on.

There is a hard boundary here, and he puts it in plain sight. The QR standard asks an encoder to score eight mask patterns and pick the least ugly. Paris fixes mask 0 and bakes it into the glyph outlines, noting that automatic mask selection "would probably exceed the Reed–Solomon portion in complexity." A machine that can only substitute can compute a remainder; it cannot compare two futures. Saying that out loud is worth more than engineering around it.

The picture is made the same way. Each bit position has two glyphs, D37_0 and D37_1: one empty, one carrying a filled square at that module's coordinates, the two swapped wherever the fixed mask bit is 1. Every intermediate glyph has zero advance width, so they all stack at a single origin. A final QR_BASE glyph supplies the finder patterns, timing patterns, format bits, quiet zone, and the only nonzero width in the run. The image, in his phrase, is "the union of hundreds of tiny glyph outlines."

The order of the pipeline leaves a wound. Browsers break lines on the raw Unicode before they call the shaper, so they cannot know the text is about to become a QR code. Chrome splits at a space or a dot and shapes each run separately; the second run has no opening bracket, so it renders, honestly enough, as coded]. There is a quieter scar too: he deleted the GDEF table outright, because Firefox subpixel-snaps zero-advance mark glyphs against their nearest base glyph and the parity section drifted sideways from the data. Part of this font is defined by an omission.

QR codes are finished construct. ISO/IEC 18004, infrastructure, and "QR art" has had a name for years. Type design is finished construct. Even the fact that OpenType shaping can be made to compute is not new — Paris mentions in passing a 2026 demonstration that implemented a raycaster inside the TrueType hinting VM. The beauty of an already-construct is sediment from a chisel-construct cycle that has already closed. It is good to look at. It has stopped growing.

What has no name yet is this object. It is not a typeface — nobody will set a book in it. It is not a program — it does not execute. It is not net art — it lives in a repository with a couple of dozen commits, made by an engineer who did not bother writing an about page. (The design memo reads like the transcript of a conversation; its citation links still carry ChatGPT's tracking parameters. That does not change what the artifact is.) The art world will eventually find a word for typefaces that do arithmetic, and on that day this gets filed. It hasn't been yet.

The reason to look now is plain. The parity bits have always been in front of you. You have looked at them every time you have scanned anything — 56 of the 208 data modules in a Version 1 symbol, saying nothing at all. Making a remainder visible usually means lifting it out and setting it on a plinth. Paris did the opposite: he built a machine that runs almost entirely for the sake of the remainder, and the remainder finally showed its size.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics