ErotemeEroteme AI
How Eroteme's 8-agent consensus system generates calibrated predictions on real-world events.
Eroteme AI is the platform's multi-agent prediction engine — an 8-agent system organised into two specialised layers, with a meta-judge that synthesises all outputs into a single calibrated probability.
Instead of relying on one model's opinion, Eroteme AI runs five different AI providers across eight distinct analytical roles, then uses extended reasoning to produce a final verdict that's more accurate than any single model.
The 8-Agent Architecture
Eroteme AI operates in two parallel layers, each with four specialised agents:
Research Layer (runs first, in parallel)
| Agent | Provider | Role |
|---|---|---|
| News Researcher | Perplexity | Real-time web search for the latest developments |
| Sentiment Analyst | Grok | X/Twitter social sentiment analysis |
| Data Analyst | Gemini | Current statistics, data, and historical precedents |
| Base Rate Analyst | Claude | Historical frequency analysis — ignores current news to avoid recency bias |
Analysis Layer (runs second, in parallel)
| Agent | Provider | Role |
|---|---|---|
| Bull Case | Claude | Constructs the strongest case for YES |
| Bear Case | ChatGPT | Constructs the strongest case for NO |
| Evidence Synthesiser | Gemini | Balanced, objective evidence weighting |
| Contrarian | Grok | Identifies where consensus may be wrong |
The analysis layer receives a summary of the research layer's findings, so each analyst has access to real-time data, sentiment, statistics, and base rates.
Meta-Judge
After all 8 agents complete, Claude with Extended Thinking acts as the meta-judge. It doesn't simply average the agents' outputs — it uses Bayesian reasoning to synthesise all 8 perspectives, weighting by confidence and historical accuracy per agent.
The meta-judge produces:
- A calibrated probability (0–100%)
- A confidence interval (uncertainty bounds around the probability)
- A disagreement score (how much the agents diverge)
- Key factors driving the prediction
- What could be wrong — the biggest risk to the estimate
Two Tiers
| Standard | Eroteme | |
|---|---|---|
| Agents | 1 (single model) | 8 (full pipeline + meta-judge) |
| Real-time data | Limited | 4 research agents with live data |
| Perspectives | Single model | Bull, Bear, Evidence, Contrarian |
| Cost | $1 USDC | $5 USDC |
| Refund guarantee | Yes | Yes |
Why This Works
- Errors cancel out — One model's blind spot is covered by another
- Forced disagreement — Bull and Bear agents ensure both sides are argued
- Calibration feedback — The meta-judge receives historical accuracy data per agent and adjusts weights accordingly
- Transparency — You see every agent's individual probability and reasoning