Dear Ms. Nelson,
Digital Denials documented the failure in terms the industry has not answered: the most common landlord response to a dispute, observed by 86 percent of the attorneys and counselors you surveyed, is to ignore the dispute and reject the applicant anyway. A dispute right that terminates in a form letter is not a dispute process. The reason it terminates there is structural: neither the landlord nor the screening company holds a per-applicant record of why the denial happened, so there is nothing to dispute against.
We build and run a system that produces exactly that record, and we have published a working description of it with its evidence public. The mechanism is small: the housing provider's written screening criteria are pinned to a cryptographic hash, the application file is hashed record by record, and every disposition is derived from those criteria by several independent model seats whose reasoning is compared criterion by criterion. The system writes a disposition record the applicant can read — the criterion that fired, the record it fired on, what was absent, and what evidence would reverse the decision. Where the file does not support a conclusion, the system abstains rather than denies by default, and the abstention is itself a signed record.
The article is here: https://miscsubjects.com/a/tenant-screening-adverse-action-record
It is written against the instruments your work has called for: FCRA's adverse action and accuracy duties, HUD's 2024 holding that denials must not arrive in a conclusory fashion, and the SafeRent settlement's five-year injunction. The measured evidence is on the page — a 30-case adjudication calibration study (two model families, 30/30 and 29/30 correct dispositions, zero wrongful authorisations) and the hash-chained receipts for false convergence, parser voids, and abstentions, each linked from the article.
This letter is a permanent object. Its full text hashes to 90fa3b326e04f40cc3efc53b11877293c080eadd10b410a62dd3217892f4e13d. It was sent autonomously-written and owner-approved on 2 August 2026, is receipted on the article it belongs to, and any commercial AI model pointed at this site can explain it in full.