700 Judges, 3.2 Million Cases: The System That Decides Asylum Cases

700 Judges, 3.2 Million Cases: The System That Decides Asylum Cases

Two asylum seekers cross the southern border six months apart. Same country, same persecution, same statute. A clerk neither of them has ever met routes one to the regular docket and the other to an expedited docket. One gets heard in eight months. The other is waiting until 2028. That single routing decision is now six years of difference. In the first of a two-part deep dive on immigration adjudication, Gwen and Marc walk through the system that will decide both of their cases: * 3.2 million pending cases, 2.3 million of them asylum, divided among roughly 700 immigration judges — about 3,200 cases per judge, each requiring hours of fact-intensive testimony through interpreters about events in another country * The particular social group doctrine — the single most contested category in asylum law, which has flipped four times in eleven years through Attorney General certification. Same statute, same words, opposite results depending on who holds the office when your case is heard * How the AG cleared 767,000 cases in 2025 — the highest single-year total in the system's history — through a mix of procedural exits, compressed dockets, and doctrinal changes that foreclosed categories of claims that were viable when applicants filed * The MSPB decision that stripped immigration judges of civil service protections, holding they're inferior officers removable at will — meaning the judge deciding whether to grant asylum on a borderline theory now knows she can be fired tomorrow with no notice and no appeal * The denial rate climbing from roughly fifty percent in 2023 to eighty percent in February 2026, without a single statutory change Rosa's hearing is in 2028. She'll argue a social group theory the Attorney General has already foreclosed, with six-year-old evidence about a country she fled six years ago, in front of a judge who can be removed at will. The statute hasn't changed. Everything else has. Next episode: the person standing inside the machinery — detention, representation, expedited removal, and why the Constitution doesn't fix any of it.

The Framework That Decides Disability

Until June 2024, the Social Security Administration was denying disability claims on the grounds that applicants could work as pneumatic tube operators — a job that functionally disappeared decades ago. The agency's catalog of occupations hadn't been updated since 1991. Congress noticed. Courts complained. The agency issued two emergency messages. The fix removed 114 jobs nobody was citing anyway and added a documentation requirement to 13 jobs people were. The vocational expert who cited pneumatic tube operator last year can cite it again next year, as long as they say an extra sentence. That catalog isn't the problem. It's the symptom. In this episode, Gwen and Marc walk through what Social Security disability adjudication actually looks like when you run it two and a half million times a year: * The five-step sequential evaluation that decides every claim — substantial gainful activity, severity, the listings, past work, and the step five question that wins or loses most cases: is there any other job in the national economy you could do? * The grids — literal rows and columns where you plug in age, education, work history, and RFC, and the table tells you whether you're disabled. The Supreme Court blessed this trade in Heckler v. Campbell: uniformity over individualized judgment, explicitly because the alternative would "hinder needlessly an already overburdened agency" * The age cliff: at fifty-four, the grids say not disabled. At fifty-five, same medical condition, same RFC, same education, same work history — disabled. The borderline rule that's supposed to soften this treats five months as too far away * How every structural feature from the trilogy — the forty-six point judge-level variation, the weak internal review, the ex ante control mechanisms — runs through the credibility finding and the RFC, which determine where you land on the grids * The Matthews problem at scale: the test counts administrative burden as a factor, and at two and a half million applications a year, that factor dominates the math. There's no doctrinal floor on how thin process can get once the system is big enough People die waiting for hearings. The GAO and Social Security's own inspector general have documented it. At a nine-month average wait — an improvement from the fifteen months it was running in 2023 — that number is meaningful every year. That's not the system failing. That's the system working as designed, at volume.

Nobody Said Deny More Cases: How Agency Preferences Reach the Hearing Room

In the final episode of a three-part series on how agencies actually produce outcomes, Gwen and Marc trace the mechanisms that did the work — all of them upstream of the hearing room and mostly invisible from outside: * Case completion targets that measure speed but not thoroughness — and a Seventh Circuit concurrence warning that even well-intentioned production pressure "can alter the essential function of adjudication" * Quality review rubrics that aren't published, shift with administrations, and can't be appealed — but feed into performance evaluations that have real career consequences * Social Security Rulings that describe themselves as not having "the force and effect of law" and in the same sentence say they are "binding on all components" — binding policy issued without notice and comment * The 1984 case that saw it coming: Association of Administrative Law Judges v. Heckler, where the court found outcome-based targeting of ALJs "violated the spirit of the APA, if no specific provision thereof" * Immigration as the amplified version: 700-case-per-year quotas for judges without APA independence protections, deciding cases where the stakes are deportation Somewhere out there is a claimant with the same diagnosis as someone approved three years ago by the same ALJ in the same office. That claimant gets denied — not because the case changed, but because the system around the judge changed. And nobody outside the agency can see why.

The Review Layer Doesn't Fix It — Three Things Agency Review Does Instead

In 2018, Attorney General Jeff Sessions reached into a single immigration case, certified it to himself, and overruled the precedent that had let domestic violence survivors qualify for asylum. In 2021, Garland certified it back. In 2025, Bondy certified it away again. Three reversals on the same doctrine in seven years — no statute changed, no rulemaking happened. One signature each time. In this episode — the second in a three-part series on how agencies actually produce outcomes — Gwen and Marc move up a level from the hearing room and ask whether internal appellate review fixes the variation E13 documented. The textbook answer is yes: that's what review is for. The real answer is that "internal review" describes three very different machines, and none is primarily in the business of producing consistency. * The APA baseline is permissive. 5 USC 557 gives the reviewing body "all the powers which it would have in making the initial decision" — including the power to redo credibility findings from a cold transcript. The ALJ who sat three feet from the witness gets reversed; the federal court then defers to the agency under substantial evidence review. The one person who actually saw the testimony is the one whose findings get displaced. * Three agencies, three pullbacks. Social Security walks the default back via 20 CFR 404.970 — the Appeals Council acts only on legal error or want of substantial evidence. Immigration walks it back furthest via 8 CFR 1003.1 — IJ fact findings get clear-error review. The SEC operates closest to the default, where "all the powers" actually bites. * Model 1 — Social Security: review mostly doesn't happen. The Appeals Council grants review in roughly 15-20% of cases. The rest get a one-paragraph denial. That denial opens the 60-day window to federal court, but reads to a lay claimant like a final rejection — because in ordinary English "denied" means you lost, while in admin-law English it means you've exhausted and the clock is running. Represented claimants know this. Unrepresented claimants miss the window and lose by default. * Model 2 — Immigration: review is a policy instrument. The BIA designates decisions as precedential, binding every IJ in the country. Under 8 CFR 1003.1(h), the Attorney General can pull any case — pending or decided — and write a precedential opinion on it. Matter of A-R-C-G (2014), Matter of A-B- (2018), Garland's vacatur (2021), Matter of S-F-M (2025): the same doctrine swung four times across four administrations through a single referral mechanism. That isn't error correction. It's substantive policymaking through the review channel. * Model 3 — Enforcement (SEC, FTC, NLRB): review is automatic and institutionally tilted. Every ALJ decision goes up. About three-quarters of SEC reversals favor the enforcement division — finding violations the ALJ didn't or raising penalties the ALJ imposed. The Commission supervises the enforcement staff, sets priorities, allocates their budget, and then sits as the appellate tribunal reviewing whether that same staff proved its case. No bad faith required; the incentives only run one way. Three failure modes, one conclusion: SSA fails by not reviewing; immigration fails by reviewing through a politically directed mechanism whose priorities shift with administrations; enforcement fails by reviewing through a body with an institutional stake in the result. Review corrects some errors and generates precedent where precedent gets designated. What it does not do — anywhere — is pull similar cases toward similar outcomes. Next episode: agencies don't shape outcomes by correcting decisions. They shape the conditions under which decisions get made.

The Forty-Six Point Spread: Why Your Outcome Depends on Who You Get, Who Represents You, and Where You Live

A 2017 GAO study found that Social Security disability outcomes could swing by forty-six percentage points based solely on which ALJ heard the case. Same claimant, same record, same hearing office, different judge. Individual ALJ approval rates range from under ten percent to over ninety. That's not noise — it's the system. In this episode — the first of a three-part series on how agencies actually produce outcomes — Gwen and Marc step outside the hearing room and examine the three variables that shape results before a single question gets asked: * Judge assignment: Random allocation protects against selective routing, but paired with forty-six points of variation, it produces something that looks like a lottery. The system chose neutrality over consistency — and consistency doesn't get recovered downstream. * Representation: Claimants with attorneys get roughly double the approval rate of those without — not mainly because of what happens in the hearing, but because of record-building beforehand and something more structural: since ALJ decisions aren't published, the only way to know how a specific judge handles specific issues is to have appeared in front of them. Experienced local attorneys hold what amounts to privatized law — accumulated operational knowledge that doesn't exist in any public source. * Geography: Hearing offices differ by ten to fifteen points in approval rates beyond what the mix of judges explains, driven by local economic conditions, regional medical infrastructure, and office-level practice cultures that develop over time. A national program, a single statutory standard, and systematically different outcomes depending on which judge you draw, whether you can afford a lawyer who knows that judge, and which office covers your zip code. The binding agency-level law that does exist — Social Security Rulings, HALLEX — covers a narrow band of interpretive questions. Everything outside that band is where the variation lives. Next episode: if horizontal consistency doesn't exist at the hearing level, does top-down appellate review fix it? Spoiler — it doesn't.