Workshop · floor model for Survey of Software 1.080.1

qrkadelic

Plain language. Switch to Engineering for versions, byte counts and the measurements behind each claim.

Your payload, four ways

Four separate libraries — written by different people, in different decades — encode whatever you typed above. Here is what each one produced. Look at them before reading on.

Encoder parity: qrcode, segno and qrcodegen encode the same payload at the same error-correction level. Rendered side by side, then compared module by module.

qrcode — when it has to look like something

The code itself is fixed, so all the fun is in the drawing. Punch a hole through the middle and see how big you can get away with.

Centred occlusion sweep against error-correction level. Decoding verified live in-page; the reader's phone is the second instrument.

The surprise: most of the benefit arrives at Q. Going from M to Q buys you a lot of room; going on to H buys almost none — the code mostly just gets bigger.

But only if something is actually covering the code. Every number on this page measures how much can sit on top of a symbol. If nothing does — no logo, no penguin, just a plain code on a page — the trade runs the other way. Higher error correction is paid for in extra squares, and squares are what a camera runs out of first. Cover the code: pick Q. Leave it alone: pick M, and spend what you saved on printing it bigger.

THE CEILINGS GOVERN OCCLUSION AND NOTHING ELSE. They answer "how much can sit on top of this symbol." They do not answer "can a camera resolve this symbol," and for a clean code the two constraints pull in opposite directions: error correction buys damage tolerance by spending modules, and module count is the binding constraint on a distant, tilted or dim capture. Choosing Q for a symbol nothing occludes buys recovery that will never be used and pays for it in density.

From the field, not measured in this tab — this is not hypothetical. These figures were carried into a live tool with no logo on its code, Q was chosen citing them, and a Pixel read the page as text rather than finding a symbol. On that tool's payload — a whole trail packed into a URL, far longer than the one above — Q gave 125 modules across and M gave 105. Under a simulated camera — rendered at real display size, then downscaled, blurred, tilted and noised — Q at 150px failed below 400 sensor pixels while M at 340px read at 300. Right numbers, wrong question.

Measured on this payload with the penguin at this aspect and pad, as the widest occluder that still decodes — L 13.5%, M 18.5%, Q 28.0%, H 29.5% of symbol width (2.5%, 4.7%, 10.7%, 11.9% by area). Same units as the slider.

M→Q is the jump; Q→H is nearly flat. Both Q and H force version 4, so the extra ECC codewords are largely consumed by the larger symbol — with a centred square occluder the two are identical at 11% of area; with the penguin's shape and pad, H buys about 1.5 points of width over Q. Either way, the widely-repeated "30% of a QR code can be damaged" is codeword recovery capacity and not area you may paint over.

These ceilings are not a decoder artefact. Swept across all four zxing-cpp binarizers — LocalAverage, GlobalHistogram, FixedThreshold, BoolCast — the limits are identical, because covered modules are destroyed data rather than data that is merely hard to read. Confirmed on hardware: at every error-correction level, a Pixel gave up at the same cutoff this page did. That is the sharp contrast with colour, where the same phone read a bright tie-dye at EC=L that this page had called unreadable — colour failures are a statement about the scanner, occlusion failures are a statement about the code.

segno — when it has to print, and stay small

Same code, written out as a vector for a printing press. The file sizes are not close — but it depends which drawing method you ask for.

SVG byte counts, generated live. Three bars rather than two: the headline 16× describes qrcode's default per-module emitter, and its SvgPathImage factory closes most of that gap.

pyStrich — more than QR, without a compiler

Not every barcode is a QR code. Same payload, other symbologies.

Pure-Python multi-symbology encoder: Data Matrix, PDF417, Aztec, Code128 and more, plus DXF output for engraving.

qrcodegen — the same encoder in six languages

This one draws nothing at all. It hands you the grid and stops — which sounds useless until you need the identical encoder in Java, Rust and TypeScript too.

Matrix-only reference implementation, one source across Java/TS/Python/Rust/C++/C. 0.04 MB installed, zero dependencies, no renderer of any kind.

zxing-cpp — real, but not runnable here

PyQRCode — the empty space

Start your build

Take everything you just worked out and hand it to whichever AI you use. This copies your settings, the limits you found by dragging, and working code for exactly this configuration.

Clipboard envelope — no vendor deep link, no API key, no dependency on anyone's roadmap. Paste into any assistant.