Single questions have a price; copying it is hard to beat. "A and B" has no price. The model must read the prices and reason about how A and B depend on each other.
Setup
The model sees all of Manifold as of 7 days before close: every market, price, trade and comment. Marginals come from the market, so coherence holds by construction.
Bar to beat
339 test pairs where both markets were open 7 days before close. Dashed line: prices with A and B assumed independent.
Steps
1Manifold snapshot at any date, leak-tested: building.
2More pairs where both markets are open at the cutoff, grouped by type: running.
3Qwen3-235B and GPT-6 Luna with search tools, zero-shot.
4Train Qwen3-0.6B only if step 3 beats prices + independence.