Independent research lab
Open research. Inspectable systems.
Pre-registered research on design systems and accessibility. Hypotheses are locked before estimation, every figure traces to a versioned artifact, nulls and deviations are reported, and the data can be queried in your browser.
Published paper
Do design systems deliver accessibility at scale?
Strong design-system adoption signals were associated with roughly half as many automatically detected accessibility violations. The pre-registered model replicated on held-out UK data.
- US association
- −50%
- Held-out UK replication
- −44%
- Sites analysed
- 18,547
What I built
The study is also a software system: collection, transformation, estimation, publication, and exploration share versioned contracts.
Scanner
Playwright + axe-core collection with frozen detection and politeness rules.
ETL
DuckDB pipelines that emit compact Parquet and typed JSON artifacts.
Analysis
Pre-registered PPML models, diagnostics, replication, and simulation tests.
Visualisation
Accessible, responsive figures shared between paper and plain-English article.
Explorer
DuckDB-WASM queries over versioned datasets, entirely inside the browser.
Reproducibility
Public preregistration, data freeze, deviations, source, and generated results.
Data explorers
Each dataset is a versioned Parquet artifact queried locally with DuckDB-WASM. Queries and filter choices never leave the browser.
US federal websites: USWDS adoption x accessibility
One row per live federal .gov/.mil website with a completed accessibility scan: graded US Web Design System adoption signals (uswds_count, semantic version) alongside axe-core violation counts in 11 categories, plus agency and digital-maturity covariates. Derived from the GSA Site Scanning daily snapshot by tools/etl.
Explore dataUK public sector: govuk-frontend adoption x accessibility
One row per UK public-sector website with a completed homepage scan: a graded govuk-frontend detection score (structural mirror of GSA's uswds_count), axe-core violation counts in the same 11 categories, organisation type (central, local, parish, NHS, devolved), and HTTPS/page-weight covariates. Collected for Paper 01 by tools/scanner under a frozen recipe.
Explore dataResearch discipline
Variables, models, and priors are locked before confirmatory relationships are examined; exploratory findings are labelled; null results are reported as nulls; and replication on held-out data is designed in from the start. The approach bounds researcher degrees of freedom without pretending observational data can prove causation.