Sophecs

Philosophy and AI, taught through allegiance. A ten-question diagnostic places you in one of six schools of ethics (Stoicism, Utilitarianism, Virtue Ethics, Confucianism, Ubuntu and Buddhist Ethics), as a point on a map of the six. You then argue an AI-related motion from that school, and a judge scores how faithfully you argued from it rather than whether it agrees with you. A lessons library holds the primary-source passages behind each motion.

Live at sophecs.com. Built for 15- to 18-year-olds.

The six are shown by the SIX_SCHOOLS=1 flag, which production has had since the launch on 2026-11-08. A build without it is the site as it was before: three schools and the first quiz. docs/stage5-plan.md has the history.

Running it

npm install
npm run dev

The site builds and renders without any environment variables, but the parts backed by a service degrade rather than work:

• No Supabase credentials: quiz submission fails gracefully and /r/[id] returns 404. Sign-in, the arena, and /me need a real project.
• No ANTHROPIC_API_KEY: /api/judge returns its paused state instead of a verdict.
• No NEXT_PUBLIC_POSTHOG_KEY: analytics calls no-op.

.env.example documents every variable. supabase/migrations/ holds the schema, applied in order. npm run seed:topics upserts content/topics/ into the debate_topics table; the other content directories are read off disk at build or request time and need no seed step.

Checks, all five of which must pass before a push:

npx tsc --noEmit && npm run lint && npm run test && npm run build && npm run check:bundle

npm run check:bundle reads a build without the flag and fails if one of the three new schools' id, name, adherent noun or colour is in any client chunk, and npm run check:bundle:six checks that a flag-on build carries them. CONTRIBUTING.md says why, and docs/review-charter.md is the standard a step is reviewed against.

Judging is live, deliberately

KILL_SWITCH_JUDGE=true is still the default in .env.example, and while it is on no argument is sent to the judge and the debate screens say so. Arthur has turned it off in production, having reviewed a batch of real verdicts for quality himself. He said so directly, and that instruction is the only record: the production value is not in this repository and no commit here changed it. docs/decisions.md carries the entry.

Do not change the switch or JUDGE_ALLOWLIST_USER_IDS. Real submissions are being scored, so the standing rule is that no production verdict may change: any change to what the judge receives needs Arthur's yes, and proof that the request is byte-identical where a fixture covers it: tests/judge-request/six-schools/ for the six on a flag-on build, and tests/judge-request/ for the three on a build without it. Unchanged fixtures do not by themselves clear this. docs/review-charter.md rule 2 says what else counts.

The judge itself is a real model call, not a stub: src/lib/anthropic.ts sends the prompt in content/prompts/judge.v3.md on a flag-on build, as production is (judge.v2.md on a build without the flag), to claude-sonnet-5. Each school's fidelity criteria are in src/lib/judge/rubric.ts. It scores three criteria, in this order of weight: Fidelity, Rigor and Engagement. It returns a combined score out of 100 alongside them. The criteria, and what the score does and does not mean, are published at /how-it-works (on a build without the flag, /debate/rubric and /method).

Where things live

• content/topics/: six motions, frontmatter only, one file each. Each carries a one-sentence stance for all six schools; motions are not owned by a single school.
• content/micro/: twelve micro-lessons, a before and an after for every motion, each with one primary source, and six "view from elsewhere" passages, one per motion. Read by src/lib/micro-lessons.ts.
• content/schools/: the six school pages behind /s/[school].
• docs/quiz-v2.json: the ten diagnostic questions and the map, as Arthur accepted them; content/quiz/questions.ts is the first quiz, which a build without the flag still asks.
• content/prompts/: the judge prompts (judge.v1.md, judge.v2.md and judge.v3.md, the one production sends) and the counterpart screening prompt.
• content/modules/: six longer teaching units, one per school, plus a README documenting the schema. /lessons grows a Modules section from them. It was empty when this section was written.
• src/lib/supabase/: the Supabase clients. There is no fixtures fallback; without credentials the backed features degrade as listed above.
• src/lib/elo.ts: Elo, K=32.
• supabase/: schema and row-level security policies.

Design system

Paper and ink; colour is allegiance. The six school colours appear only to mark a school, and the quiz result card is the single fully saturated surface in the product. Spectral carries anything that is an argument, IBM Plex Sans carries the interface, IBM Plex Mono carries anything measured. Tokens are defined once in src/app/globals.css under Tailwind v4's @theme static. No dark mode. The written plan is in design/tokens.md and design/layout.md.

Not built

No real-time features, notifications, payments, or moderation queue beyond the counterpart reply screening. There is no evidence yet that any of this improves anyone's reasoning, and the site makes no such claim. A further school is mostly a content change, since everything is keyed by school id; the three added at the launch also needed their criteria approved and a place on quiz v2's map.

Contributor rules, including the frozen paths and the review loop, are in CONTRIBUTING.md. The standard a step of work is reviewed against is docs/review-charter.md, and the decisions already taken and why are in docs/decisions.md.