What

A minimal web app under demos/typescript/ demonstrating the published @learning-commons/evaluators SDK, structured as two features so more evaluators can be added later:

1. Demo app scaffold — a runnable Vite + React + Express shell: a home page that lists evaluator demos, client-side routing, and a backend that each evaluator plugs into.
2. Math Standards Alignment evaluator page — the first evaluator. Pick a grade (K–12) and jurisdiction → searchable multi-select of that grade/jurisdiction's math standards → enter a word problem → Evaluate → view the raw alignment JSON.

Design

• Stack: React (Vite) + react-router-dom frontend; small Express backend. The SDK is Node-only (top-level node:* imports on the eval path), so all SDK calls run server-side and the browser never sees API keys (Vite proxies /api to the backend).
• Extensible by construction: evaluator demos are declared in one registry (src/evaluators.tsx) that both the home page and the router consume; the backend exposes GET /api/health and a single call site where each feature registers its routes. Adding an evaluator = a new page + one registry entry + a new 003-… spec, touching nothing else.
• Math feature endpoints: GET /api/jurisdictions (from the SDK's Jurisdiction enum), GET /api/standards (Knowledge Graph lookup), POST /api/evaluate (MathStandardsAlignmentEvaluator.evaluateItems). If the required keys are absent the routes still mount but return 503, so the scaffold and other evaluators keep working.
• Server lifecycle: graceful SIGINT/SIGTERM shutdown (server.close + closeAllConnections) so dev-server restarts don't hang or orphan the process.
• Spec-driven artifacts per feature: specs/001-demo-app-scaffold/ and specs/002-math-standards-alignment/, each with spec.md / plan.md / tasks.md.

SDK gaps surfaced (follow-ups, out of scope here)

• No public standards-listing helper. KnowledgeGraphClient isn't exported, so server/kg.ts calls the KG REST API directly to populate the picker. A future release should export a getStandardsByGrade-style helper.
• Peer range too loose. The SDK declares ai >=6.0.0 but is built against ai@7 (@ai-sdk/provider spec v4); installing ai@6 with @ai-sdk/anthropic@4 throws a runtime "Unsupported model version" error. The demo pins ai@^7; the SDK peer should be tightened to ai >=7.

Verification

• tsc --noEmit clean; vite build clean.
• Standards lookup verified live against the KG; /api/health and route wiring smoke-tested.
• Real-key end-to-end run verified via full UI click-through (grade → jurisdiction → standards → word problem → Evaluate → result).
• Findings from a self code-review and two rounds of Copilot review addressed: stale-response race guard, KG pagination + missing-UUID guards, backend-down error handling, and input validation (including blank statement codes).