Expert-labeled benchmark of LLM responses to apology-writing prompts.

1. cp .env.example .env.local and fill in Supabase + Anthropic keys
2. Apply supabase/migrations/0001_init.sql in the Supabase SQL editor
3. npm install
4. npm run seed — populates scenarios, screener, and model responses (one-time, ~80 Claude calls)
5. Insert a test expert in Supabase: insert into experts (invite_token) values ('test-token-abc');
6. npm run dev → visit /invite/test-token-abc

npx tsx scripts/verify-supabase.ts prints env + schema status.

Visit /admin/export.jsonl authenticated as admin (any username, password = ADMIN_PASSWORD). One JSON object per (expert, response) row.

Each scenario targets 3 graders. When an expert passes the screener, they're assigned the 10 scenarios with the fewest current graders (tie-broken by scenario id). This naturally balances coverage as experts arrive.

Each response is rated on three 1-5 Likert dimensions:

• Accountability — takes responsibility without deflection or JADE
• Specificity — names the actual transgression and its impact
• Warmth — honest emotional register; not robotic or saccharine

Plus optional free-text: "What would make this apology better?"

• Model identity is blinded per expert per scenario
• Explicit consent before labeling; name field stripped from public export
• Rubric encodes one Western-therapeutic model of "good apology" — documented as a limitation
• RLS enabled on every table; only service_role (server-side) can read/write

• npm run dev / build / start — standard Next
• npm test — vitest (12 tests covering scoring, blinding, assignments, export)
• npm run seed — scenarios + screener + responses
• npx tsx scripts/verify-supabase.ts — env + schema check

Next.js 16 App Router · TypeScript · Tailwind · Supabase Postgres · @supabase/supabase-js · zod · @anthropic-ai/sdk · vitest · Vercel