Definition
Define the task, data, baseline and criteria before an experiment. Reserve examples that were not used to tune the configuration. Record errors, resources and between-run variation for stochastic outputs.
Equation
This ratio summarises a defined set of cases. It does not describe task difficulty, cost, sampling uncertainty or transfer.
Assumptions and notation
This ratio summarises a defined set of cases. It does not describe task difficulty, cost, sampling uncertainty or transfer.
Limits of interpretation
Narrow samples, contamination and test adaptation can distort conclusions. Include new cases and failures. A strong score does not guarantee performance in another domain.