The Physical Cost of Useful AI Work
Task-normalized energy accounting for multi-step and agentic AI workflows.
Research thesis
Energy per token is useful for serving diagnostics, but it is not the physical cost of useful work. Critical Load measures the energy required to obtain an accepted task outcome, including failed attempts, retries and orchestration within the declared boundary.
What the package contributes
A frozen metric system, task registry, run-log schema, NVML measurement wrapper, metric calculator, release validator, open evidence register and an external-measurement worked example.
Use class
Suitable for
Benchmark design; model/workflow routing; engineering regression; vendor measurement requirements; energy-aware capacity planning.
Not yet suitable for
Critical Load model rankings; universal agent-efficiency claims; facility energy without facility measurement; claims that primary telemetry has already been collected.
Release gate to v1.0
Execute repeated primary runs on declared hardware, publish raw attempt-level telemetry, audit acceptance scoring, reproduce every table from raw logs and complete independent systems/energy and domain review.
Critical Load Research Group - September 2026 - CL-WP-005 v0.9