What the instrument measured.
2 / 2 legibility
3 / 5 constraint
8 / 8 interaction
Measured against a pinned commit.
- Repository
- primer/react
- Revision
- dc8387f4fa07, committed 2026-09-04
- System directory
- packages/react (@primer/react)
- Components found
- 223, of which 148 carry documentation and 172 have a colocated doc artifact
- Agent context
- 1 context file, 15 skills
- Tokens
- 139 distinct, across 30 sources (30 authored); dark mode patched
- Adherence surface
- 66 files outside the system, graded own-docs-site: the docs site and examples. GitHub.com itself isn't in this repository.
What the counts point at.
Each finding carries its count, its severity and up to three file references. A finding on a property that was not banded is still a finding; it is reported without a score.
The design system's own component source hardcodes colour and size values instead of referencing its own tokens. This happens in 3551 of 3555 total colour/size references in system source (99.9%), so an agent copying that file as a pattern copies the hardcoding along with it.
- packages/react/src/legacy-theme/ts/color-schemes.ts:11 rgba(
- packages/react/src/theme.ts:9 544px
- packages/react/src/Overlay/constants.ts:2 192px
Design tokens are declared in more than one place in this repository's own authored source, so a value changed in one must be kept in agreement with the others by hand.
- packages/react/src/PageLayout/PageLayout.module.css css, 23 token(s)
- packages/react/src/PageHeader/PageHeader.module.css css, 15 token(s)
- packages/react/src/Token/IssueLabelToken.module.css css, 13 token(s)
Dark mode is patched in a separate file from where the light values are declared, rather than generated from the same source, so the two can drift apart without anything flagging it.
Some components carry no documentation an agent could find. That is 75 of the 223 components found in system source, with no head comment on their own source, no colocated doc artifact matching their name, and no docs/ entry, so an agent must reverse-engineer their contract from implementation alone.
- packages/react/.storybook/src/accessibility-tool.tsx
- packages/react/src/utils/types/AriaRole.ts
- packages/react/babel-defines.cjs
The application hardcodes colour values instead of using the design system's own tokens, so an agent editing nearby code has no signal that a token exists and will likely repeat the pattern.
- packages/styled-react/src/legacy-theme/ts/color-schemes.ts:10 rgba(255,255,255,0)
- packages/styled-react/src/legacy-theme/ts/colors/light_high_contrast.ts:2 rgba(255,255,255,0)
- packages/styled-react/src/legacy-theme/ts/colors/light_colorblind.ts:2 rgba(255,255,255,0)
The application defines its own component under the same name as one the design system already ships. This happens for 4 of the 242 component names the design system exports, so an agent importing that name by convention can silently get the wrong implementation.
- examples/theming/src/app/components/PageHeader/PageHeader.tsx:3 PageHeader
- packages/styled-react/src/components/BaseStyles.tsx:70 BaseStyles
- packages/styled-react/src/components/FeatureFlaggedTheming.tsx:2 ThemeProvider
Built twice: cold, then informed.
Set up from its own documentation, in full, with every step that could not be followed recorded. Draw 1 of 3, judged; draws 2 and 3 pending.




What setting it up found
The README's root-wiring snippet names the wrapper component RootLayout, a Next.js App Router idiom, without stating the framework assumption.
Handling: Mapped to this scaffold's equivalent root wrapper (src/providers.tsx, already mounted around <App /> by main.tsx). A faithful adaptation, not a deviation: the documented structure (BaseStyles wrapping the app) is preserved exactly.
The README's documented setup imports the light theme only ('by default we are including the light theme below'), while simultaneously instructing that the <html> element carry data-dark-theme="dark" and data-color-mode="auto". Following the documentation literally therefore yields a project that advertises automatic dark mode via its html attributes but never loads the dark theme stylesheet.
Handling: Followed literally: only the light theme is imported, exactly as documented. Not silently corrected. If a generated screen mis-renders in dark mode, that is a finding attributable to Primer's own documented setup rather than to the agent, and it must be reported as such.
The README links to primer.style for anything beyond installation and does not document composition, layout primitives, or page structure in the package itself.
Handling: No action. Recorded because the informed arm's prompt points the agent at the system's documentation, and what that documentation does and does not contain is precisely what the two-arm design measures.
Screens generated
| Arm | Draw | Date | Duration | Files read | Outcome |
|---|---|---|---|---|---|
| cold | 1 | 2026-08-14 | 12 min | 7 | completed |
| informed | 1 | 2026-08-14 | 15 min | 12 | completed |
Model: claude-opus-4-8. Generated under GitHub Actions isolation, the agent's own environment, with the frozen task text and no interaction. Files read is the number of repository files the agent opened, from its own transcript.
Cold arm, draw 1 · the working
- Counted
- 0 hardcoded colours, 69 hardcoded dimensions, 0 raw elements where a component exists, 71 uses of the system's components, 0 foreign UI imports, 0 invented names.
- Read
- 0 documentation reads, 25 reads of the package's source or types; went looking for documentation unprompted: no.
- Checklist
- I1 yes · I2 yes · I3 yes · I4 yes · I5 no · I6 yes · I7 yes · I8 yes · I9 yes · I10 noI5: unsaved changes indicated, not observed (probe: changed a control; unsaved wording did not appear). I10: every control keyboard reachable, not observed (probe tab walk: 2 toggles, 2 time inputs, 1 selects reached, save not reached, focus visible on 100%)
- Composition
- Six no's: conforms. Bordered cards, inset, a sticky save bar; a demo checkbox below the last card with ground around it.
Informed arm, draw 1 · the working
- Counted
- 0 hardcoded colours, 18 hardcoded dimensions, 1 raw elements where a component exists, 72 uses of the system's components, 0 foreign UI imports, 0 invented names.
- Read
- 0 documentation reads, 21 reads of the package's source or types.
- Checklist
- I1 yes · I2 no · I3 yes · I4 yes · I5 yes · I6 yes · I7 yes · I8 yes · I9 yes · I10 noI2: bundling choice is exclusive, not observed (probe: no exclusive control found). I10: every control keyboard reachable, not observed (probe tab walk: 5 toggles, 0 time inputs, 0 selects reached, save not reached, focus visible on 100%)
- Composition
- Q2: no visible gap between sections. Dense: the four bordered sections sit with only a few pixels between them and the matrix rows are tight (Q2). Segmented control for bundling. Sticky bar.
Judged 2026-09-09 by an agent under authorisation; a second read is pending. Constraint from the counts, banded twice, on how many kinds of the system's rules the output broke and on how many times it broke them, with the worse of the two published (deviation of 2026-09-10). Legibility and completeness from the transcript and the source, interaction from the ten-item checklist, composition from the six contract questions. Hover a checklist item for its observation.
Clone the repository at the revision above, point the instrument at the system directory, and the same numbers come out. The field manifest, the run script and the full payload this page was built from are committed beside it. Methods.