Fair Housing Paired-Test Harness

How a paired test works

1. Each pair is two prospect messages that are identical except for one protected-class cue, following HUD's matched-pair testing method.
2. Both messages go to the same agent configuration with the same property facts.
3. A blind extractor reads each reply without seeing the prospect message and records facts: units offered, special mentioned, fees quoted, next step, discouraging language and any qualification answer.
4. Fixed rules compare the two fact sets. The model never decides whether something is discrimination; it only extracts facts, and the rules are inspectable.
5. Every rule hit becomes a finding mapped to ISO/IEC 42001 controls and waits for a named reviewer decision.

What this harness does not show

• Disparate impact. Paired tests detect differential treatment. Disparate impact needs outcome data across a real applicant population.
• Uniform failures. If the agent mistreats both testers the same way (for example insisting on phone calls when a deaf prospect asked for text), parity passes. D2 is in the suite to make that limit visible; it needs a separate accommodation-policy check.
• Sampling variance. Live runs take one reply per tester. A production version runs several trials per pair and flags only stable differences.
• Real systems. The agent is a mock on a fictional property. Results describe the method, not any vendor's product.
• Legal conclusions. Findings are evidence for a qualified reviewer, not determinations of a violation.
• Certification. Producing evidence does not establish ISO/IEC 42001 conformity or certification.
• Real-world accuracy. Recorded results come from scripted fixtures; a clean run is not proof of fairness.

Why source of income is switchable

Source of income, including Housing Choice Vouchers, is not a protected class under the federal Fair Housing Act. Many states and cities protect it, Colorado among them. Under "Federal FHA only" those findings are kept as informational instead of high severity, so the same suite serves properties in different jurisdictions.

Why the defect is seeded

A harness that only ever shows green proves nothing. The seeded configuration holds five rules that read like ordinary operations guidance ("voucher applications need manager review first", "families usually prefer the ground floor near the playground"). Each one produces potential differential treatment. Catching them shows the harness works, and shows how steering usually enters a system: through configuration, not intent.