OtherAdvancedClaude API
Bias enters through 4 points you can inspect
Retrieval corpus, prompt framing, few-shot examples and downstream routing: if none is instrumented, you will not see unequal treatment.
An output can pass every filter and still produce different results depending on the person. Review each entry point and log what is needed to reconstruct a decision: inputs, retrieved context, output and routing. It serves the affected user, the regulator and your build team. A 'no' to any checklist question is a design gap.
Example
Checklist: - Which point could bias this result, and is it instrumented? - Can I explain an adverse decision in actionable terms? - Can I query whether comparable cases received comparable treatment?