Project 03Active
Apollo Predictions
About thirty AI agents that research and forecast prediction markets.
Overview
Apollo Predictions analyzes, forecasts, and resolves prediction markets with about thirty agents. Specialists cover sentiment, markets, politics, and insider signals, and an adversarial reviewer challenges the result before anything is published.
The hard part is making a long, paid, multi-model pipeline trustworthy. A run has to survive retries without repeating finished work, never publish half a result, and keep every number traceable to evidence.
Key systems
- 01
Staged research pipeline
A run moves through decomposition, parallel research, analysis, synthesis, adversarial review, and revision before anything is published.
- 02
Durable checkpoints
Each stage saves a checkpoint, so a retry resumes from completed work instead of starting over.
- 03
Numbers from fields, not prose
Every number on the dashboard is a structured field, never text extracted from a model's answer.
- 04
One publication gate
Results are published through a single validated step, so a failed check is a failed run, not a publish-and-fix.