# Critical Load Useful AI Work Measurement (UAWM) Method v0.9

## Objective
Measure operational energy against a user-valued outcome rather than an invocation artifact.

## Reporting unit
**Energy per Accepted Task (EAT)** = total measured energy across all attempts / number of accepted tasks.

An attempt is one execution toward a frozen task. An accepted task satisfies a predeclared acceptance rule. Failed attempts and retries remain in the numerator.

## Required companion metrics
- Attempt success rate
- Retry energy share
- Orchestration Overhead Index (matched workflows only)
- Energy per response and energy per token as diagnostics
- Latency
- Tool-call count
- Quality-energy Pareto frontier

## Measurement boundaries
- B0: GPU only
- B1: CPU + GPU
- B2: node
- B3: facility

Every published energy value must carry its boundary. A narrower measured boundary cannot be relabelled as a wider one.

## Primary measurement implementation
Preferred NVIDIA GPU measurement uses the cumulative NVML total-energy counter before and after each attempt. Sampled power is retained as a diagnostic channel. CPU/DRAM/node measurements are added only when actually available.

## Minimum release requirements
1. Frozen task inputs and acceptance criteria.
2. Frozen hardware/software manifest.
3. Raw attempt-level energy records including failures.
4. At least 3 independent runs per core task/model cell; more for stochastic workflows.
5. Scoring audit and documented exclusions.
6. Clean reproduction of summary outputs from raw logs.
7. Independent measurement and task-domain review for v1.0.

## Interpretation
This method is designed for model routing, workflow engineering, procurement requirements and energy-aware capacity planning. It does not create a universal energy factor for a model or agent architecture.