Random noise gates, identical protocol (val threshold pick -> test sim):
| Variant | Trades | PF | WR |
|---|---|---|---|
| ungated | 32 | 0.917 | 56.2% |
| sham_seed1 | 15 | 1.457 | 60.0% |
| sham_seed2 | 19 | 0.825 | 57.9% |
| sham_seed3 | 16 | 3.348 | 75.0% |
| sham_seed7 | 17 | 2.118 | 64.7% |
| sham_seed11 | 12 | 2.103 | 66.7% |
| sham_seed42 | 13 | 2.599 | 69.2% |
Verdict
GATE IS NOISE if random gates hit PF >= 1.3 anywhere near as often as the real-model gates (LSTM 1.99, TRM 1.69, XGB 1.27).