Anti-Discrimination Check¶
Screening test — instantiates Symmetry-Based Fairness
Holds a case fixed and flips only a protected characteristic — race, sex, religion, disability, age — to see whether treatment moves; a targeted symmetry test for the markers the law and ethics forbid from counting.
Anti-Discrimination Check is the narrow, high-stakes special case of symmetry: the irrelevant transformation it probes is specifically a protected or prohibited marker. It takes an otherwise-identical case and flips one sensitive attribute — race, sex, religion, disability, age, national origin — and asks whether the decision or its terms change. Its defining move is that the marker is declared off-limits by prior commitment: unlike a general relevant-difference test, this check pre-commits that the protected characteristic is not a legitimate basis for different treatment (barring narrow, named exceptions), so any movement under the swap is a finding, not a judgment call. And because the stakes are legal and moral rather than merely operational, a positive finding is routed to a defined review path rather than filed as a note.
Example¶
A lender wants to know whether its underwriting model treats applicants fairly with respect to race. An Anti-Discrimination Check constructs matched applicant pairs that are identical on every legitimate factor — income, existing debt, credit score, loan-to-value, employment history — and differ only on a proxy for the protected characteristic (say, applying names or neighborhoods strongly associated with a particular group). It runs each pair through the decision and pricing engine and watches the output. If two applicants who are indistinguishable on creditworthiness receive different approvals or different interest rates when the only thing that changed was the protected marker, that is disparate treatment[n1] — the marker is doing work the lender swore it would not. The check flags the pattern and hands it to the fair-lending review path, which can pull the model, investigate the driving feature, and answer to regulators, rather than letting the finding evaporate into a dashboard.
How it works¶
- Fix the protected markers in scope. Decide which characteristics are off-limits for this decision, grounded in law and policy.
- Build the counterfactual. Hold every legitimate factor constant and vary only the protected marker — through matched real-world pairs (paired testing) or a model-based swap on the attribute.
- Observe the treatment delta. Any change in decision or terms attributable to the swap is the signal.
- Route a finding. A confirmed disparity triggers a pre-defined escalation and review, not a silent log — the check's teeth are in what happens next.
Tuning parameters¶
- Markers in scope — which characteristics are tested. Each added marker widens protection but multiplies the pairs to construct and the carve-outs to reason about.
- Proxy handling — whether you swap only the direct attribute or also its correlates (ZIP code, surname, school). Ignoring proxies is the most common way a discriminatory system passes a naive check.
- Method — live paired testing versus a statistical counterfactual over historical decisions. Paired testing is vivid and legally potent but expensive; the statistical route scales but depends on model access.
- Exception allowance — how narrowly bona-fide carve-outs are drawn (a genuine occupational or actuarial justification). Loose carve-outs reopen the door the check was built to close.
- Escalation sensitivity — how large or systematic a disparity must be before the review path fires.
When it helps, and when it misleads¶
Its strength is that it is targeted, legally grounded, and actionable: it isolates the one class of asymmetry that law and ethics treat as categorically forbidden and gives a positive finding somewhere to go.
Its central failure mode is proxies. A system can discriminate through features correlated with a protected marker while sailing through a swap of the marker itself — the classic misuse is testing only the obvious attribute, finding no direct effect, and declaring the model clean while ZIP code quietly carries the same signal. It also addresses only disparate treatment: a facially neutral rule that produces skewed outcomes without keying on any protected marker will pass this check entirely, because that is a disparate-impact problem measured at the population level, not a single-case swap. The guarding discipline is to test proxies as well as direct markers, and to pair this check with a population-level distributional review. That population view is what a distributional fairness review supplies — where Classification Fairness Review measures error and outcome rates across whole groups, this check flips a single protected marker on one otherwise-identical case; they catch different failures and neither substitutes for the other.
How it implements the components¶
Anti-Discrimination Check realizes the protected-marker face of the archetype — the components that turn a general symmetry principle into a prohibited-basis guarantee:
equivalence_rule— it defines cases as equivalent when they differ only by the protected marker, pinning down the exact equivalence the check enforces.value_priority_statement— it encodes non-discrimination on the prohibited basis as the governing priority: the marker is declared, in advance, never a legitimate relevant difference (save narrow carve-outs).appeal_or_review_path— a positive finding routes to a defined escalation and review rather than a silent note, which is what gives the check consequence.
It does not fix the invariant payout or standard for equivalent cases in general (equal_treatment_rule) — that is [Policy Symmetry Test] and [Equal-Treatment Checklist]; nor does it measure realized variation across a population (audit_trail) — that is [Consistency Audit].
Related¶
- Instantiates: Symmetry-Based Fairness — this check is the archetype narrowed to the markers that are prohibited from ever counting.
- Sibling mechanisms: Policy Symmetry Test · Consistency Audit · Equal-Treatment Checklist · Precedent Analysis · Classification Fairness Review · Exception Register
Editorial Notes¶
Form Classification¶
Form family: Experiment, Test & Rehearsal
Rationale: Holds a case fixed and flips only a protected characteristic — race, sex, religion, disability, age — to see whether treatment moves; a targeted symmetry test for the markers the law and ethics forbid from counting, making its operative form a deliberate probe, variation, simulation, or practiced execution used to generate evidence or readiness.
Independent corroboration: The frozen evidence defines Anti-Discrimination Check as 'Holds a case fixed and flips only a protected characteristic — race, sex, religion, disability, age — to see whether treatment moves; a targeted symmetry test for the markers the law and ethics forbid from counting', so its operative form is Experiment, Test & Rehearsal.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Law & Governance
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Antidiscrimination law defines protected characteristics and disparate treatment, making the otherwise-identical protected-attribute swap a legally meaningful test.
Related originating lineages:
- Gender Studies & Queer Theory — Intersectional theory informs which protected dimensions and structural effects require scrutiny.
- Philosophy — Moral equality and symmetry under irrelevant differences provide the normative basis.
- Statistics & Experimental Design — Matched-pair testing and causal counterfactuals supply the empirical method.
- Ethics of Technology & AI Governance — Algorithmic fairness audits generalize the test to automated systems.
Review resolution: Antidiscrimination law is the agreed primary lineage. Equality theory, paired testing, protected-class scholarship, and algorithmic governance materially shape the counterfactual swap check; the practice is established rather than Encyclopedia-created.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Disparate treatment in U.S. antidiscrimination law is treating someone worse because of a protected characteristic — the mirror image of disparate impact, where a facially neutral rule produces unequal outcomes without keying on the marker. An anti-discrimination check targets the former: change only the protected attribute and see whether treatment moves. ↩