A browser experiment can measure worker duration, owned WASM memory, GPU buffer sizes, structural unit counts, cache bytes, and task outcomes. These are useful inputs to an internal resource ledger.
Do not overname the units
Elapsed milliseconds are affected by scheduling. Allocated bytes are not total device pressure. A battery percentage is not per-action energy. Unless a proxy is independently calibrated, describe it as an algorithmic cost or normalized budget rather than joules.
Keep safety external
Even a teleodynamic controller needs hard ceilings for memory, latency, and user cancellation. The experiment should record those guards separately from the endogenous ledger.