A design system that judges its own work.

Agents write most of DEED’s interface. A style guide a model reads once is advice it can forget, so I wrote the system as rules a machine can run. Everything that exists is in a catalogue: 229 components, each with its states, its properties and a quality score. If it isn’t in the catalogue, an agent can’t use it.

The rules are short. Tokens only, never a raw value. One acid-yellow signal per screen. Amber, the colour of agents and the operator, on at most 7% of it. Two-pixel corners. Motion only when something changes state. A number that hasn’t arrived shows a placeholder, never a made-up figure.

A critic measures thirteen rules on the rendered page, not the source, at 390, 768 and 1280 pixels: one focal point per screen, the amber budget, motion with a purpose, contrast, fit. Thirty-six pages are checked and four are exempt, each with a written reason. Every pass is tied to the exact file it judged, so changing a page voids it.

Twenty-three checks run in a chain before anything counts as done, and the first one tests the others: each must still catch a defect planted for it, or the chain is not trusted. Every number in the docs comes from one machine source, so a stale count fails the build.

The mark is a cursor, and its bar is the only part that moves. Motion has three verbs: settle, land and refuse. Nothing bounces. Sound is the same cursor heard: one root at 110 hertz, whole-number harmonics only, off until you turn it on.

Type is Söhne in four cuts. Weights step between eight licensed masters and are never interpolated.

A manual in thirteen volumes, from capital and governance to memory and accessibility, shows every component at work. Each chart answers one question, inside a frame that names its source and how often it updates.

An outside agent can build with the system without reading its code: a summary written for language models, an MCP server over the catalogue and the rules, a command-line tool and a package registry.

On 31 July 2026 I gave a fresh agent an empty folder and nothing but those public docs. It built a screen that passed every check: catalogued components only, one signal, honest numbers, no raw colours. It also found nine gaps. The docs taught the rules well and the locations badly: the index said the command-line tool did not exist when it did.

Version 3.171 shipped on 4 August 2026, the 281st numbered release since 1.0 on 29 May. Its own scorecard gives it 84 out of 100, and 29 of the 229 components are still below its quality bar.

The checks held; the words around them drifted. The stranger test and the scorecard found the same weakness: the rules were taught well, and the documentation fell behind the code. The repository is private.