Recommended records
- Data: dataset version, split method, authorisation boundary and exceptional samples.
- Model: code version, configuration, weights, random seed and training log.
- Evaluation: metric definitions, thresholds, script version and sample scope.
- Error analysis: representative failure cases, cause categories and hypotheses to verify.
- Deployment: runtime environment, dependency versions, inputs, outputs and resource use.
Acceptance criteria must be defined for the actual task and sample conditions. Performance outside the recorded test scope is not implied.