Adversarial search engine. 4 AI agents research, attack, and judge answers before showing them to you.
Site: atlantiskb.com
You ask a question. The system searches the web, builds an evidence pack, and runs it through a pipeline:
- Researcher drafts claims from the evidence
- Adversary attacks those claims
- Critic evaluates the attacks (in Strict mode)
- Judge rules under 6 hard rules — no fabrication, no speculation, no health/legal advice without flagging it, cite partisan sources, require source diversity
Every citation gets verified against the actual source. Confidence is scored from features, not vibes. The full audit trail is visible.
- Fast — 3 agents, ~30s, simple factual questions
- Strict — 4 agents, ~60s, contested or complex topics
- Liability — 4 agents with safety caps, health/legal/financial queries
The system picks the mode automatically.
git clone https://github.com/teddygcodes/atlantis.git
cd atlantis
pip install -r sydyn/requirements.txt
export ANTHROPIC_API_KEY=your-key
export TAVILY_API_KEY=your-key
# Ask a question
python3 -m sydyn "What is the speed of light?"
# Run the API server
uvicorn sydyn.api:app --host 0.0.0.0 --port 8000Python, FastAPI, Next.js, Tailwind, Anthropic Claude API, Tavily Search API, SQLite, Vercel, Railway.
The Atlantis engine (V1–V2) was the research system that proved the architecture — 20 AI States across 10 domains generating and attacking hypotheses under constitutional governance. Sydyn is the search product built on top of that work.
MIT