Applied Form engine

Standards by Applied Form

A design-system conformance gate measured in a real browser. Give it a DTCG token file and a page; it renders the page in Chromium at four fixed widths and two sampled ones, under reduced motion and in print, and checks every computed value against the system's rules. It reports what the browser drew, not what the stylesheet says.

Standards is open so the evidence about it can be checked. Every comparison it is quoted in can be rerun from this repository: the gate beside axe-core and stylelint, the drift survey of public design systems, and the model study and its judging rounds, with every record they were scored from.

Start

npm install @appliedform/standards && npx playwright install chromium
npx standards init --pack instrument  # load a system: the free Standards pack, or your pack's tokens.json
npx standards lint page.html          # your page, judged by that system's rules, at every width

With no system loaded, standards lint judges a page by the rules that hold for any page (contrast, focus, target size, overflow, alt text, heading order, motion) and says how to load one. For an agent, npx standards mcp-config prints the server block for Claude Code or Cursor.

Use

npx standards lint page.html https://example.com/page      # every rule, every width
npx standards lint https://example.com/ --register ii      # a page that has not declared a register
npx standards baseline https://example.com/a --out standards-baseline.json
npx standards lint https://example.com/a --baseline standards-baseline.json
npx standards fix page.html                                # apply the fixes it can make; writes page.fixed.html
npx standards audit pages/ --fix                           # every page in a folder, as found and as the fixer would leave it, as one report
npx standards drift https://example.com/ --out drift.json  # what system a site actually renders
npx standards rules                                        # every rule, its severity and what it checks
npx standards mcp                                          # the same tools for a model, over MCP

standards fix exits 0 when the fixed page passes every checked rule, 1 when violations remain, and lists what it could not fix by rule and element. standards audit writes one self-contained HTML page (and --json) with the findings by rule and by page, the WCAG criteria the rules touch, what is not checked, and the command that repeats the run. standards lint exits 1 on any error-severity violation not held in the baseline. standards.config.json sets a rule to error, warn or off; standards.config.example.json is a starting point. The exception is a mandatory rule: contrast, text-too-small, target-size, focus-hidden, link-not-underlined, alt-missing, horizontal-overflow and wordmark-live-text, listed under $extensions["form.applied"].levels.mandatory in the token file. A config that sets one of them to warn or off is refused, and standards lint exits 2 (decision 0027 in the Applied Form repository). A system loaded over the token file may raise the thresholds contrast and text-too-small read (floors.body-contrast, display-contrast, large-text-px, large-bold-text-px, text-min-px) and is refused if it lowers one below the token file's own value. A system that declares a pack and states no register.pack is refused too: it judged every page by register alone and then dropped that house rule, so no mandatory rule was measured. So is a system with a register whose background is not a colour. Four gaps are known. On a page whose register the loaded system does not define, with or without a system, a config that lowers register leaves the mandatory rules unmeasured. A system file that declares itself a pack is judged without text-too-small, link-not-underlined and wordmark-live-text. On a page that paints no ground, contrast measures against the ground the system names. The MCP tools standards_lint and standards_fix measure only the widths a caller names, so a call without 320 never measures horizontal-overflow at 320px.

Names that changed

The product was called the engine, and its command af, in 3.x. The old names still run in 4.x, each says once on stderr that it is deprecated, and they are removed in Standards 5.0.0. Uninstall @appliedform/engine before installing @appliedform/standards: both ship the af and af-mcp bins, so a global install of one over the other fails with EEXIST.

3.x	Now	The old name
package @appliedform/engine	@appliedform/standards	Uninstall @appliedform/engine first
command af	standards	Runs the same command after one line on stderr
af-mcp, af mcp	standards mcp	Serve as before, listing the af_* tools
tools af_lint and the eight others	standards_lint and the eight others	Both names are accepted by every route
af.config.json	standards.config.json	Read, with one line; the new file wins if both exist
af-baseline.json	standards-baseline.json	Read and written at its own path, with one line
.af/	.standards/	A path to it in a config is a plain path and keeps working
AF_BRAND, AF_CHROMIUM	STANDARDS_SYSTEM, STANDARDS_CHROMIUM	Read on every use, with one line; both names of a pair set to different values exits 2
af-audit.html, the default report file	standards-audit.html	None
the engine bin	none	None

Documents recorded before the rename, such as the decision records and the evaluation protocol and its results, use the old names, and are not rewritten.

The token file

tokens/applied-form.tokens.json is the Reference system: Standards' structure (registers, voices, spacing, floors and every rule's parameters) with a neutral palette derived by standards brand. It exists so Standards can be tested and published. It is not a design to ship. Point it at your own system by replacing the file, or by importing a Standards pack with standards system import.

Evidence

What	Where	Rerun
The gate beside axe-core and stylelint on planted defects	eval/compare.mjs, test/compare.test.mjs	npm run eval:compare
Drift across public design systems	eval/drift/	npm run drift:batch -- all
Models generating pages with and without the rules	eval/, docs/evaluation-protocol.md	npm run eval -- --models stub offline, then with keys
A judging round with anchors and attention checks	eval/judge-run.mjs	npm run eval:judge

The protocol was written before the runs it governs, and it says what result would count against Standards. These run from a clone of this repository: the npm package carries Standards and nothing that needs a model SDK, axe-core or stylelint.

Develop

npm install
npm run check    # build, unit tests, and the clean fixture linted in a real browser

Tests use node:test. Set STANDARDS_CHROMIUM to a Chromium executable if Playwright has not downloaded one.

Licence

Apache-2.0. See LICENSE and NOTICE. The name Applied Form and its marks are not licensed.