Gated architecture · input 10 features · top-3 gate on layer 1 · remaining 7 join every layer after · log-scaled parameter axis · counts measured from real model builds
| Layers | MAX params (256u + BN) | MIN params (16u, no BN) | Note |
| 1 | 2,312 | 88 | |
| 2 | 70,920 | 472 | |
| 3 | 139,528 | 856 | |
| 4 | 208,136 | 1,240 | |
| 5 | 276,744 | 1,624 | |
| 6 | 345,352 | 2,008 | current cap |
| 7* | 413,960 | — | if cap lifted |
| 8* | 482,568 | — | if cap lifted |
| 9* | 551,176 | — | if cap lifted |
| 10* | 619,784 | — | if cap lifted |
* rows 7–10 are extrapolated: the genome pool currently caps architectures at
6 layers × 256 units (mutation can only add layers up to that limit). If that cap were
lifted, the ceiling keeps climbing linearly (~68,608 params per extra 256-unit layer).
Beyond the data: on CPU (this machine) a ~1M-param model takes ~48s per 10-epoch eval vs
<1s for the current best — so the cap is a practical guardrail, not a hard limit.