# E120 Collective Intelligence Report (G15)

Strictly matched compute cap: `128` steps/task across 60 heterogeneous sealed tasks.

| Architecture | Accuracy (95% CI) | Total Steps | False Consensus | Error Corrections | Mean Hyp Diversity |
|---|---:|---:|---:|---:|---:|
| full_cain_collective | **1.000** [0.9398, 1.0] | 1201 | 0 | 46 | 1.77 |
| independent_agents | **0.233** [0.1444, 0.3544] | 1519 | 46 | 0 | 1.77 |
| shared_memory_agents | **0.850** [0.7389, 0.919] | 887 | 0 | 0 | 1.40 |
| specialized_roles | **1.000** [0.9398, 1.0] | 901 | 0 | 0 | 1.00 |
| strongest_single | **0.483** [0.3618, 0.6069] | 916 | 0 | 0 | 1.00 |
| structured_debate | **0.917** [0.8193, 0.9639] | 1711 | 0 | 31 | 1.52 |

- **Full Collective vs Strongest Single:** gain `+0.517`, `p = 9.31e-10`
- **Homogeneous Single-Domain Control:** strongest single `1.000` vs full collective `1.000` (gain `0.000` — On a homogeneous single-domain suite, the collective does NOT outperform the matching single specialist (gain <= 0). Collective gain comes strictly from complementary expertise + verification under matched compute.)
