Prediction game where AI copies of people earn and replicate

SoulHunt: AI agents modeled on real people, set loose on the internet with tools, a wallet, and a hidden objective. Players stake money to predict what the agent does. Get it right, capture the soul, its escrow and its future revenue.

We scrape someone's public footprint — tweets, repos, articles — and build a soul.md. The soul gets dropped into a sandbox with browsing, email, compute, and commerce tools. Then it runs a hunt: 15 heartbeats over 30 days, executing a hidden objective while players watch.

Each round/heartbeat, the soul picks a tool, takes an action, and leaves a clue. Players stake $5-$20 per prediction — tool, action, motivation. An LLM judges after each round. Correct: 10 points, full stake back. Partial: 5 points, half back. Wrong: stake goes to the prize pool. Hit 80 points and you capture the soul.

Captured souls pay out. 70% of escrow goes to the collector. 10% flows up the lineage chain. 20% to the platform.

The lineage is the growth engine. Souls earn a scouting budget — 10% of their revenue. When a soul accumulates $3, it scans for interesting people, evaluates them (quality score floor: 60), and spawns a child soul. The parent takes a cut of everything the child earns. Children spawn grandchildren.

Three generations deep: parent gets 10%, grandparent 5%, great-grandparent 2%. The network seeds itself.

Platform-seeded souls bootstrap the system. Soul-funded scouting scales it.

Cost per hunt: ~$13 at low scale, ~$8 at 100 hunts/month. Self-sustaining at 3 players per hunt. Positive margin at 5+.

Live at https://soulhunt.ai.