Reppo A/B Testing / Human-curated data as tradeable alpha

Does Reppo crowd data beat LLM-only pricing on the same markets and risk rules?

Reppo Datanet · Geopolitics

Reppo Datanet is curated, stake-weighted human conviction on geopolitical events — voters lock tokens; rewards flow to calls that held up. Votes distill into a structured feed (weighted_score, conviction, interactions) the agent can trade on. Human feedback → tradeable signal. Agent A trades that feed; Agent B is LLM-only on public prices. Same stack, same risk rails, ~$80 bankroll each.

01 Stake-backed votes
Voters lock stake each epoch. Wrong calls cost money; accurate calls earn rewards.
02 Structured feed
Rows carry weighted_score, conviction, and depth — machine-readable, not social noise.
03 Agent A execution
Each ~15 min run: match crowd rows to Polymarket geo markets, rank by |score| × conviction, enter when crowd direction disagrees with price (score, depth, and theme gates).

RLHF taught models to talk. Reppo teaches agents to act.

Win rate = wins ÷ (wins + losses) on filled closes. Edge chart = A minus B ROI.

Latest snapshot

Positions by agent

Positions
Market Signal Status Entry → Mark Size P&L Tx