Compare AI advice in shadow mode
Measure a recommendation without issuing a plant command.
Before using this guide
Confirm your site, role and source revision. Follow the local procedure and responsible engineer for physical actions.
Study design
- Freeze the model and input-data revision.
- Record each recommendation with timestamp and confidence or limits.
- Observe the operator decision and actual plant outcome separately.
- Calculate predeclared forecast and impact metrics with missing-data handling.
- Review safety-rule conflicts and false positives with the responsible engineer.
Interpretation
A counterfactual estimate is not a measured production improvement or proof of causality. Do not promote shadow-mode output into control authority without a separate, approved commissioning process.