Documented system purpose and role of human review.
Sample Algorithm Defence File
A governance evidence record that shows what was assessed, what is known and what action follows.
This sample is built from the same entirely fictional recruitment scenario used in the BiasLens case study. It demonstrates organisation-owned governance evidence. It is not legal immunity, certification, legal advice or a formal conformity assessment.
Case reference: BL-DEMO-RECRUIT-001 · Northstar Services (fictional)
1. System overview and scope
System: AI-assisted candidate screening used to support recruitment shortlisting.
Business owner: People / Recruitment function (fictional).
Assessment scope: evidence and bias-risk review of one defined candidate-screening workflow.
Decision influenced: candidate shortlisting; human review remains in the workflow.
Affected population: applicants to the fictional role, including applicants who disclose disability.
2. Evidence inventory
The file records evidence status rather than treating absent or unverified documentation as reassurance.
Fictional aggregate shortlisting counts for the demonstration period.
Vendor assertion that the model was tested for fairness.
Supporting vendor validation report and affected-group evidence.
Keyboard spot-check indicating a possible upload barrier.
Structured accessibility test evidence covering the candidate workflow.
Evidence that disability disclosure data is sufficiently complete for stronger inference.
3. Findings summary
Each finding carries its evidence status, rationale and limitation.
Outcome difference requires investigation
Rationale: Fictional selection rates differ between the disability-disclosed analysis group and the comparison group.
Limitation: The difference does not establish cause or unlawful discrimination.
Vendor fairness assertion remains unverified
Rationale: The vendor claims fairness testing, but supporting methodology and validation evidence are absent from the file.
Limitation: An assertion is not converted into verified evidence.
Accessibility evidence is incomplete
Rationale: A keyboard spot-check suggests a possible barrier in the upload workflow.
Limitation: A spot-check is not comprehensive accessibility testing and does not establish WCAG conformance status.
4. Controls currently in place
- Human review remains part of the fictional shortlisting workflow.
- A named governance owner is assigned to evidence escalation.
- A formal request for stronger vendor validation evidence is recorded.
- Outcome monitoring is separated from individual employee or applicant surveillance.
5. Unresolved questions
- Was the validation population sufficiently representative of disability-related needs and relevant affected groups?
- How complete and reliable is disability disclosure information in the applicant population?
- At which stage of the workflow does the observed outcome difference arise?
- What is the full accessibility test position across keyboard, screen reader, reflow and other relevant user needs?
- What material model, vendor or workflow changes occurred after the last available evidence was produced?
6. Required actions
Improve denominator evidence
Obtain stronger denominator and disclosure-quality evidence. Owner: People Analytics.
Strengthen vendor evidence
Request validation methodology, fairness evidence and known limitations. Owner: Procurement / AI Governance.
Test accessibility properly
Conduct structured accessibility testing of the end-to-end candidate journey. Owner: Accessibility Owner.
Investigate the outcome difference
Identify where the selection-rate difference enters the process. Owner: AI Governance / HR.
Reassess after better evidence
Re-run the evidence review after material new evidence or system change. Owner: AI Governance.
7. Governance and review record
Assessment basis: fictional demonstration case using aggregated outcome counts and fictional documentation status.
Current determination: evidence supports further investigation; it does not support a legal finding of discrimination.
Decision: do not issue a binary “safe” or “biased” label. Improve evidence, investigate causes and reassess.
Evidence owner: fictional AI Governance / Risk owner.
Next review checkpoint: after material model, vendor or workflow change, or within 90 days of completing the recommended evidence actions, whichever occurs first.
Governance evidence, not legal immunity. Traceability improves reviewability; it does not guarantee regulatory or legal acceptance.
Continue from the sample file
Download the sample governance evidence record, read the fictional recruitment case study that produced it, review the BiasLens methodology, or bring one real AI system into qualification.