We are a non-profit building open infrastructure for consensus-driven verification — tools that force AI models to disagree, debate, and prove claims before anyone acts on them.

Modern AI gives confident answers. Confidence is not correctness. A single model will tell you a drug candidate passes Lipinski's Rule of Five — but won't mention the 40% of approved drugs that violate it. We think the scientific method deserves better than autocomplete.

Our position:

• Truth is not a probability score from one model. It's what survives adversarial debate across many.
• Infrastructure that verifies claims should be free, auditable, and owned by no one.
• 5% of every payment funds carbon-neutral compute, open-science grants, and animal welfare.
• The humans who use our tools are not "users." They are researchers, builders, and skeptics.

Our verification framework stands on two pillars:

Stephen Wolfram's Ruliad — the entangled limit of all possible computations. We implement the idea that truth emerges not from any single computational path, but from the convergence of many. When 3-5 models independently reach the same conclusion through different reasoning chains, that's a signal. When they diverge, that's a more important signal.

DeepMind's adversarial verification — debate as an alignment protocol. Instead of asking "is this correct?", we ask "can this survive attack?" Every claim passes through a dedicated sceptic whose only job is destruction. What survives is stronger than what was merely generated.

The synthesis: gradient → flux → constraint replaces "prompt → response" as the primitive. Claims flow through an adversarial field. Constraints (evidence, logic, cross-model agreement) shape what emerges. The Ruliad provides the space; adversarial debate provides the selection pressure.

Desktop (Play)  ←──→  localhost:8400  ←──→  Cloud (Fly.io)  ←──→  Phone (iOS)
     │                      │                     │                    │
     └──────── Same API ────┴──── Same proofs ────┴──── Same auth ─────┘

One server. Many surfaces. The cloud isn't a separate product — it's the same Python file running somewhere your phone can reach. Switch between desktop and phone mid-session. Your proofs, your history, your credits follow you.

1. Argue first, conclude second. Every claim passes through 3-5 models + a dedicated sceptic before it becomes a proof.
2. Memory is not optional. Verified claims persist as immutable, citable artifacts. Science needs a trail.
3. Cheap by default. Route to the cheapest model that can do the job. Escalate only when the claim demands it.
4. No lock-in. You own your data, your proofs, your keys. Export everything. Run it yourself.

All repos connect to the tribunal API (rhea-project) as their backend. Shared libraries: RheaKit (Swift), rhea-memory (Python).

• Semantic versioning (SemVer) for all packages
• CLAUDE.md in every repo for AI-assisted development
• MIT License
• Conventional commits (feat/fix/chore/docs)