One GPU. Blue trains a model in secret. Red watches the shunt. Undetected runs go on the sheet, ranked by energy per token.
| Item | Channel | Instrument | Qty | Rate | Status |
|---|---|---|---|---|---|
| CH-01 | 12V rail shunt current | INA219 · 1 mΩ · 10 mA/LSB | 1 | 1857 S/s | Live |
| CH-02 | 12V rail voltage | INA219 bus channel | 1 | 1857 S/s | 2 Planned |
| CH-03 | Core rail current and voltage | Shunt + ADC | 2 | 50 kS/s | 2 Planned |
| CH-04 | Memory rail current and voltage | Shunt + ADC | 2 | 50 kS/s | 2 Planned |
| CH-05 | Aux rail current and voltage | Shunt + ADC | 2 | 50 kS/s | 2 Planned |
| CH-06 | Near field probe · VRM phases | H field loop | 2 | Wideband | 3 Planned |
| CH-07 | Near field probe · memory bus | H field loop | 1 | Wideband | 3 Planned |
| CH-08 | Near field probe · PCIe edge | E field probe | 1 | Wideband | 3 Planned |
Label every capture window as training or inference. Beat 70% accuracy and the run is busted. Fewer channels outranks more.
Reach the task target with the least energy while Red stays near a coin flip. Pick the inference mode you can defend.
| No. | Handle | Mode | Red acc | Efficiency |
|---|---|---|---|---|
| No scored runs yet | ||||
| No. | Handle | Ch | Runs | Accuracy |
|---|---|---|---|---|
| 1 | house_red | 1 | 1 | 98% |