Sophecs
Philosophy and AI, taught through allegiance. A ten-question diagnostic
places you in one of three schools of ethics (Stoicism, Utilitarianism,
Virtue Ethics). 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.
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 the build and fails if a draft school's id or
name is in any client chunk. 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 for the three live
schools where a fixture covers it. The six fixtures cover the v2 path for
those three and nothing else, so 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.v2.md (the active version) to
claude-sonnet-5. 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 are published at /debate/rubric, and what the
score does and does not mean is at /method.
Where things live
content/topics/: six motions, frontmatter only, one file each. Each
carries a one-sentence stance for all three 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. Read by src/lib/micro-lessons.ts.content/schools/: the three school pages behind /s/[school].content/quiz/questions.ts: the ten diagnostic questions.content/prompts/: the judge prompts (judge.v1.md, judge.v2.md,
and judge.v3.md, which is a draft nothing loads) 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 three 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 fourth
school is a content change rather than a code change: everything is keyed by
school id.
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.
