Ongoing evaluation: purpose
Ongoing evaluation looks for drift and for changes that affect bias during operation and decommissioning.
What clause 9.3.1 asks
The second procedure watches for drift and for changes that affect bias while the system runs and while it is retired. Watching continuously is how drift is found. Drift can start in the data sources, in society or in a big change to the setting. Retirement is part of the watched life, not a point where watching stops.
Why it matters
The clause makes monitoring a bias activity, not only a performance activity. A model can keep its accuracy while its decisions become less fair, because the population has changed or because the relationship between an input and an outcome has moved. The social source of drift is easy to miss in a technical dashboard. The decommissioning point is a reminder that the wind-down can affect people who depend on the system.
What good looks like
Build the monitoring plan around the bias questions, not only the model metrics. Include data sources, decisions, stakeholder effects, risks and impacts, complacency, feedback loops, the user interface, culture and the effect of deployment on external bias. Set the review points for operation and for decommissioning. Name the person who reads the results and the route for a finding.
Evidence an assessor asks for
The assessor will ask what would be noticed if the population changed. A plan that monitors only accuracy will not notice a fairness drift.
- Monitoring scope The bias questions and the items monitored, including social and cultural signals
- Operations plan The review points during operation and the named reader
- Decommissioning plan The monitoring and review during wind-down
- Finding route The route from a monitoring finding to an evaluation and a decision
Where it commonly fails
- Monitoring is a model accuracy dashboard
- Social and cultural change is not monitored
- Monitoring stops at the decommissioning decision
What to do before publication
Add the bias questions to the existing monitoring rather than building a separate dashboard.
Read across
- 5.2.5 Evaluation in the life cycle
- 9.1 Purpose
- 9.3.2 Ongoing evaluation: inputs
- 9.3.4.1 Set up the ongoing procedure
- E Cultural Aspects
← 9.2.5 Evaluation of design and outputs: outcomes9.3.2 Ongoing evaluation: inputs →