Compound Fork — Recession × Compute scale
Each cell = a joint scenario "Both branches fire". Cell intensity = total |Δ| across the 200 highest-conviction tradeable predictions vs their unconditional posterior. Joint probabilities approximated via log-odds combination (assumes conditional independence given the prediction — coarse but bounded). Click a cell to drill into the dominant single-fork view.
Pick two fork families
| Recession ↓ × Compute scale → | COMPUTE_1GW_2027 prior 60% | COMPUTE_10GW_2028 prior 40% | COMPUTE_100GW_2030 prior 20% | COMPUTE_STARGATE_FAILURE prior 15% |
|---|---|---|---|---|
RECESSION_2026 prior 20% | 135 claims · Σ|Δ| 18.91 | 140 claims · Σ|Δ| 19.83 | 138 claims · Σ|Δ| 20.41 | 134 claims · Σ|Δ| 18.76 |
RECESSION_2027 prior 30% | 135 claims · Σ|Δ| 18.93 | 141 claims · Σ|Δ| 19.90 | 139 claims · Σ|Δ| 20.48 | 134 claims · Σ|Δ| 18.79 |
RECESSION_2028 prior 30% | 135 claims · Σ|Δ| 18.93 | 141 claims · Σ|Δ| 19.90 | 139 claims · Σ|Δ| 20.48 | 134 claims · Σ|Δ| 18.79 |
NO_RECESSION_5Y prior 20% | 136 claims · Σ|Δ| 18.99 | 142 claims · Σ|Δ| 19.95 | 140 claims · Σ|Δ| 20.53 | 135 claims · Σ|Δ| 18.85 |
Method note
Joint conditional probability is approximated via log-odds combination: logit(P(pred|A,B)) ≈ logit(P(pred|A)) + logit(P(pred|B)) − logit(P(pred)). This is the closed-form Bayesian update assuming A and B are conditionally independent given the prediction. It's correct when the two scenarios act on the prediction through different causal paths; it's pessimistic when they overlap. The exact joint requires running the Gibbs sampler with both scenarios clamped, which would be N×M=16 sampling runs (~12 minutes per refresh) instead of N+M=8 — a 2× cost for higher fidelity.