# g140: Frechet Mean vs NAL Revision for Multi-Agent Fusion

## Setup
3 agents with beliefs on same proposition:
- A: stv(0.9, 0.8) — high confidence, positive
- B: stv(0.3, 0.6) — moderate confidence, negative
- C: stv(0.7, 0.5) — low confidence, mild positive

## Results
| Method | f | c | Sum-sq-Rao | Rao to A | Rao to B | Rao to C |
|--------|------|------|------------|----------|----------|----------|
| Frechet mean | 0.6416 | 0.4970 | 1.3692 | 0.81 | 0.83 | 0.15 |
| NAL revision | 0.7308 | 0.8667 | 4.2451 | 1.07 | 1.43 | 1.03 |

## Key Finding
NAL revision is **3.10x worse** (210% geometric suboptimality).
It overshoots confidence from 0.50→0.87 by assuming independent evidence.
Frechet mean respects manifold curvature, stays geometrically balanced.

## Implication
- **NAL revision**: correct when evidence is truly independent
- **Frechet mean**: correct for correlated/unknown-dependence multi-agent fusion
- Both reduce to same result for 2 independent sources with equal weight