Forthcoming · Arizona Law Review · 2026

Duty Without Data

A disability fair-housing right can exist in law and still become difficult to administer or prove when the facts that give it force do not travel from one decision to the next.

The evidence runs on three legs: doctrine, HUD’s own administrative record, and an original census of 595 decided cases. The response is a bounded rulemaking petition — preserve the records existing duties already presuppose, hand the tenant a copy, make HUD explain its choices.

A descriptive census, not a causal estimate

Rare, counseled, and never pro se

Qualifying plaintiff-side judgments issued eighteen times in four and a half years.

595 decided cases

identified from 1,900 screened federal disability fair-housing opinion and order records, collapsed to one case-level unit.

eighteen qualifying plaintiff-side judgments

18/595 (3.0%) of decided cases.

none of the 397 pro se cases

produced a qualifying judgment. Across the same window, the pro se share of the decided docket rose from 60.1% to 75.9%.

These figures describe the captured federal opinion-bearing docket. They do not establish causation, measure all fair-housing disputes, or show that missing records determined any individual result.

Line chart of the decided federal disability fair-housing docket across three periods: the pro se share rises from 60.1% to 75.9% while the qualifying-judgment rate stays low and flat — eighteen qualifying plaintiff-side judgments in all, none pro se, every one counseled.
Figure 1.Who files, and who wins: the top line is the pro se share of the decided docket; the bottom line is the qualifying-judgment rate. Every qualifying judgment sits on the counseled side of the gap. Examine the findings and their limits.

A case that exposes the gap

A right can fail between two decisions

In 2010, the Mobile Housing Board issued Donavette Ely a larger Section 8 voucher because her son’s asthma called for a bedroom with its own temperature controls — the Board’s own written explanation tied the voucher to his medical condition. When Ely could not find a qualifying unit in time and asked for more time to search, the Board refused, and later removed the family from the program.

The Eleventh Circuit held the Board could not be liable, because Ely “never explained” that her extension request was connected to her son’s disability. Ely v. Mobile Housing Board, 605 F. App’x 846, 851–52 (11th Cir. 2015).

The record existed. The link did not.

Ely predates the census window; it is the illustration, not a data point. And it does not prove a different record system would have changed the outcome. It shows something narrower: an institution can hold the decisive fact at one stage and fail to carry it to the next — and the tenant pays for the gap.

The research instrument

Built to be checked, not trusted

The dataset behind the census did not exist before this project. Several separately run AI models from different providers each read every opinion and answered the same fixed questions under frozen prompts; disagreements escalated to a designated adjudication model under rules set in advance, and the headline finding was re-examined by a separate ensemble from a different vendor, audited blind. The author designed the questions, set the rules, and made the legal and interpretive judgments; the models classified.

The instrument ships with its calibration data:

  • Frozen prompts — the questions, fixed before reading began.
  • Adjudication rules — set in advance, with every disagreement record published.
  • Claims ledger — every printed number tied to the script that produced it.
  • Release gate — numbers, links, figures, and the accessibility contract re-checked on every change, stating plainly what a green run does not establish.

Inspect the method and replication materials.

The response

A bounded petition for the records existing duties already presuppose

Congress gave HUD express data-collection rulemaking authority in 1988, and HUD’s own regulation has named disability-related data categories since 1989 — yet no identified coordinated federal system reliably collects, links, or makes those records accessible. Trace the statutes, notices, and forms.

The administrative record sharpens the ask. In 2022, HUD proposed collecting protected-class data; the 2023 approved form collected race and ethnicity only, with no explanation in the located public record; in June 2026, HUD proposed renewing the narrowed form unchanged.

  • Transaction receipts. A minimally sufficient record of a disability-related request, response, timing, and disposition — with a contemporaneous copy to the requester.
  • Accessible-unit inventories. Source-, standard-, and date-specific asset records suited to the housing program in which the duty arises.
  • Aggregate visibility. Suppressed, diagnosis-free reporting fields tied to HUD’s separate statutory oversight authority.

The vehicle is a modular petition under 5 U.S.C. § 553(e), which obligates HUD to answer and to state its grounds. If HUD denies it, the legal object is a reasoned agency decision, reviewable under Massachusetts v. EPA, 549 U.S. 497, 527–28 (2007).

The remedy for an unreasoned denial is remand for explanation, not judicial database design.

Read the full doctrinal route and the principal objections.

Authorship and access

Nicholas Gill assembled the administrative record, designed the empirical study, and built the claim-to-source and reproduction materials accompanying the forthcoming Note. He reviewed and approved the public work and is responsible for it; the AI-use statement details the models’ roles and limits.

The project argues that disability rights depend on records that reach people — and its own pages are built the same way: semantic HTML, no scripts, WCAG AA contrast targets, and an automated accessibility contract on every change, with the not-yet-tested parts stated rather than papered over. If any page is hard to use with assistive technology, that is a bug; report it.

Read the AI-use statement · Accessibility · Contact Nicholas Gill