Most online discussion rewards speed, volume, and certainty. I wanted to build something around a different incentive: make a clear claim, support it, answer the strongest counterargument, and update when the evidence changes.
Today I’m opening The Socratic Arena as a public beta.
👉 Try the live Arena
⭐ Explore the open-source repository
What it became
The project started as a real-time 1v1 debate app. That core still matters: server-authoritative turns, timed rounds, private arenas, audience deliberation, replays, and Elo ratings.
But a match result is not the most valuable output. The more interesting question is whether repeated performance can help someone reason better.
That led to Arena OS: one competitive identity across live debates, focused coaching, campus leagues, classrooms, and high-stakes simulations.
The current beta includes:
- A blind multi-judge panel for logic, evidence, and communication.
- Median aggregation plus judge agreement, confidence, uncertainty, and factual-claim flags.
- Longitudinal reasoning profiles across 11 skills.
- Targeted daily drills and instant AI sparring.
- Seasons, format-specific ratings, tournaments, and clubs.
- Classroom workspaces with assignments, custom rubrics, randomized positions, and AI-use policies.
- Professional simulations for negotiation, sales, design review, policy, and investor conversations.
- Appeals, moderation reports, fairness checks, and portable reasoning credentials.
- 2v2 team debates and auditable tournament advancement.
Designing AI judging for inspection, not mystique
“AI judge” is an easy phrase to oversell. The beta does not present a model as an infallible referee.
Competitive results require independent judge records. Aggregation uses the median rather than allowing one extreme score to dominate. The system preserves judge versions, agreement, confidence, uncertainty, and the evidence needed for an appeal. Tournament advancement is derived only from canonical completed matches.
Evidence matching is similarly scoped. It retrieves cited material with SSRF protection and measures claim/source support. A match is not proof of truth. It is a signal that helps a human inspect the claim.
That distinction is central to the product: AI-assisted judging, not automated certainty.
The production architecture
The system is built with React 19, Vite, Node.js, Express, Socket.IO, Supabase/Postgres, Redis, Gemini, Vercel, and Render.
Moving from an impressive demo to a public beta meant hardening the less glamorous paths:
- Durable PostgreSQL workflows and Row Level Security.
- Server-only atomic audience voting.
- Redis-backed distributed realtime state and rate limits.
- Idempotent finalization and transactional match resolution.
- Reconnect, timeout, abandoned-match, and stale-client recovery.
- Versioned PWA upgrades.
- Prometheus metrics, readiness checks, alerting, and hosted smoke tests.
- Human-labeled judge calibration and fairness pairs.
The repository includes migrations, tests, deployment validation, observability, and the evidence behind those claims.
Why public beta now?
The largest remaining risk is no longer missing feature surface. It is whether the product creates real value for real people.
The next milestone is deliberately measurable:
- 100 active public-beta users.
- 1,000 completed debates with quality and abandonment tracked.
- One recurring campus league.
- One institutional classroom or professional pilot.
- A larger calibration dataset comparing human and panel judgments.
I am especially looking for feedback on four things:
- What makes you trust or distrust an AI-assisted judge?
- Does the feedback change your next argument?
- What makes the first debate confusing?
- Is there a reason to return and practice next week?
If you debate, teach, coach, interview, negotiate, or simply want a better way to test your own reasoning, I’d love an honest trial.
👉 Enter The Socratic Arena
🐛 Report a bug
💬 Share public-beta feedback
Train your thinking. Prove your reasoning.
Top comments (0)