How it's built

Browser: Rust → WebAssembly

The exam engine, adaptive practice and every page are Rust (Velvet framework) compiled to WebAssembly. Pages are prerendered at build time, so they load and read without JavaScript; state lives in local storage and works offline (service worker).

Worker + D1

A small JavaScript Worker serves the static assets and the class API: join with a class code, passkeys or a recovery code, signed session cookies, sync of results and learning state, teacher and admin views. D1 has versioned migrations that the Worker can also apply itself.

Content as data

Each exam is a folder of JSON (exam guide, domains and weights, question bank, optional performance-based items and case studies) compiled into the engine. Adding an exam needs no code change; every count on this site is computed from the same data at build time.

On every push

1. Validate every exam bank (schema, content rules, catalog in sync, quality ratchet)
2. Test the Rust engine (cargo test) and the Worker against a throwaway local D1 (node:test), in parallel
3. Build the WASM, prerender every page, optimise with wasm-opt
4. End-to-end in headless Chromium (Playwright), desktop and mobile
5. Preview each branch on its own URL, isolated from production data
6. Deploy main, apply migrations and smoke-test that the edge serves exactly that commit

Self-hosted CI

Forgejo and Woodpecker run in rootless Podman; the agents cache toolchains, crates, npm packages and browsers in a named volume. A push is live on the edge in about two minutes. No CI minutes to pay for — the runners are our own machines.

Static checks

A checker runs on every build: broken references, bad answer keys and option counts, missing explanations and duplicate questions fail the build. Near-duplicates, letter bias and giveaway-length answers are reported, and a ratchet stops the share of “the correct answer is the longest option” from ever getting worse per exam.

Runtime statistics

Per question: correct rate, option picks, point-biserial discrimination and answer rate from exam simulations, plus student reports. Questions that look ambiguous or too easy are flagged in the admin view for review.

Honest numbers

No pass estimate is shown until there is enough evidence (20 answers, 3 per domain, a quarter of the concepts). After that, scores and pass probability are ranges with a confidence word, and every score is labelled an estimate.

Backups

Nightly export of the D1 database to our own disk (14 daily, 8 weekly, checksummed), an off-site copy (30 daily, 26 weekly) and Cloudflare's Time Travel as a second net.

Restore drill

Once a month CI restores the newest backup into a throwaway local D1 and checks the schema version and row counts against the backup's manifest.

Monitoring

A deep health check (Worker, database and the deployed commit) every 15 minutes; failed deploys, backups or drills raise a de-duplicated push alert through self-hosted ntfy.