The rules work in simulation. Now they must move onto real hardware, one step at a time:
- GPU power-limit test under real load: how much response is lost when power is lowered.
- Read-only battery reading for weeks before giving Barrel OS any command.
- Priority in the request queue: user questions before deferrable jobs, instead of power limiting alone.
- Energy per million tokens: the figure needed to justify spend, calculated from real system metrics.
- Generator set control and satellite weather forecast, for sites that have them.
We will publish every result here, including those that do not go as expected.