Engineering reference projects and assessment
Start with evidence
This is the implementation track for the S-level blueprint. S-level is our proficiency target, not an industry credential. New material is available as an explicitly unreviewed preview; automated checks do not confer human approval. Read the concepts, reproduce a failure, inspect the fix, and defend an alternative.
The eight product scenarios share a workbench so identity, persistence and protocol rules stay consistent. Separate Next.js and Cloudflare projects exercise their actual runtimes. Local fixtures deliberately avoid real credentials and paid model calls. They are educational reference implementations, not production deployments.
Learning sequence
Eight reference scenarios
| Product | Executable location | Failure exercise and reference behavior |
|---|---|---|
| Public lead funnel | labs/cloudflare-funnel and workbench lead view | Negative/replayed challenge produces no write; fresh challenge with same operation recovers receipt |
| Multi-tenant SaaS dashboard | labs/next-dashboard and workbench dashboard | Switch identity during loading; never reuse another account's data |
| Design system / partner SDK | workbench packages/ui, token compiler, two consumers | Alias cycles fail; packaged exports and styles resolve; dialog restores focus |
| Secure account portal | workbench account view | Drop response after commit; retry returns one receipt |
| Offline field app | workbench offline view | Reload retains pending IndexedDB work; conflict is not discarded |
| Data-heavy editor | workbench editor view | Undo/redo is bounded; a new edit discards redo; remote change exposes conflict |
| Source-grounded copilot | workbench copilot view | Permission and deletion filtering precede ranking; malicious text stays data |
| Agent operations console | workbench agent view | Disconnect/reconnect, stale hash, revocation and restart retain one authoritative outcome |
The editor is a bounded state/history kernel, not a full collaborative document product. The copilot is a deterministic retrieval baseline, not a evaluated production LLM. Cloudflare global cache behavior, real-provider quality, screen-reader review and learner pilots require their corresponding external/manual evidence. This distinction prevents local demonstrations from being presented as completed production validation.
Run the reference implementation
Use Node 24. From frontend-architect/labs/engineering-workbench:
npm ci
npm test
npm run build
npx playwright install chromium
npm run test:browser
npm run dev
The server listens on loopback port 4173. Select the local demonstration identity and one of the eight views. Data persists under .data during normal development. Tests use isolated or in-memory stores. Never expose this fixture identity selector publicly.
For Next.js, build the workbench UI package first, then run npm ci, npm run build, and npm test in labs/next-dashboard. For Cloudflare, use the project README to initialize local D1 and test secrets, then npm run validate and npm run dev. No deployment is required by the tests.
Assessment and reference answers
Submit a working journey, an injected failure, the failing and passing assertion, and one rejected alternative. Explain the transaction or lifetime boundary that establishes correctness. A screenshot, ADR alone, or passing text-length validator cannot substitute for runtime evidence.
Score each axis from 0 to 3: 0 absent, 1 happy path, 2 handles the declared failure with evidence, 3 defends tradeoffs and survives a changed constraint. Axes are correctness, security, user recovery/accessibility, operations, and explanation. Graduation requires at least 2 on every axis and no unauthorized side effect or silent loss of accepted work. Human reviewers should challenge the weakest assumption rather than count pattern names.
A reference answer for the lost-response exercise uses a scoped operation key, payload hash and unique database constraint; it does not allocate a new key merely because the first fetch rejected. For the cache exercise it prevents late pre-invalidation reads from republishing. For the agent exercise it reauthorizes and binds approval to the exact proposal, while distinguishing a local ledger from an external tool's idempotency contract.
Review and evidence status
The implementation manifest links every domain to source and a meaningful test file. The validation report records commands actually run. Human technical review, editorial review and a learner pilot are pending and must be recorded independently before final approval. Existing illustrative case-study metrics are not benchmark results from these applications.