Purpose
The bias profile is the through-life record of the bias work, built from data, model and code and kept across every iteration.
What clause 5.1 asks
The profile is the record that follows a system from its first design to its retirement. It holds what the bias work produces and what it concludes. The document names three places where unwanted bias can start: the data, the learned model, and the program code that makes it work. Any one of them can be the cause. Data and model can also act on each other, so a result can be skewed even after each was examined on its own.
Why it matters
The three-source view stops a team from treating bias as a data problem alone. A perfectly representative data set can still produce a biased output through a model choice or a coding defect, and a good model can be undone by a data pipeline that changes the meaning of a field. The through-life record is what lets a later reader see which source was examined and what was found. Without it, every iteration starts from memory.
What good looks like
Make the profile a real repository, not a folder of presentations. Give it an index, a version history and a named owner. For each entry, record the source it addresses: data, model, code, or a combination. When a bias is found, record where it entered as well as where it appeared, because those can be different places. Review the profile at every gate and at every trigger event.
Evidence an assessor asks for
The assessor will ask where a particular bias entered the system. A profile that records only the output metric cannot answer, and the mitigation will be a guess.
- Profile index The entries, their source (data, model, code or a combination) and their version
- Source analysis A note for each bias found, recording where it entered and where it appeared
- Owner record The named owner and the deputy for the profile
- Review log The gates and triggers at which the profile was reviewed
Where it commonly fails
- Bias is treated as a data issue only
- The profile is a slide deck that is overwritten each quarter
- No one can say which model or code change introduced a disparity
What to do before publication
Set up the repository and the source field before the first entry. Retrofitting provenance to a bias finding is usually impossible.
Read across
- 1.4 Field of application
- 5.2 The bias profile and the system life cycle
- 7.1 Purpose
- 9.2.4.2 Examine data pre-processing
- 9.3.4.2 Each iteration of the ongoing evaluation
← 4.5 Outcomes5.2 The bias profile and the system life cycle →