Appeals Test the Clerk, Never the Rule

In January 2026 a neuro-oncologist prescribed the drug Lynparza for Mason Henderson, a 21-year-old from southeastern Texas. His brain cancer was a diffuse hemispheric glioma with an H3 G34 mutation, rare enough that the World Health Organization had named it only in 2021. Cells in tumors like his carry a flaw the drug targets, and doctors had seen similar brain cancers respond to it in several earlier cases, though there was little evidence beyond that. His pharmacy benefit manager refused, and two weeks later it sent its reason: “Lynparza is not approved for the diagnosis provided.”1

The reason was accurate. Lynparza was approved in 2014 for ovarian cancer, and its label didn’t cover his. Insurers deciding whether to pay for treatment of a rare cancer generally go by FDA labels and expert guidelines. His family appealed to Jefferson County, which pays the claims for its employees’ family coverage directly, and the county’s review board sent the case to an independent medical reviewer. The reviewer, a nonspecialist, backed the denial and suggested a different drug. Henderson got Lynparza in March from AstraZeneca’s patient assistance program, which had turned his mother down two weeks earlier and shipped a bottle to her pharmacy the day after she posted about him on social media. He died at home on May 4.2

Nothing in the appeal could ask whether a rule that couldn’t see his tumor’s mutation should have been governing his treatment.

Two different failures arrive in the same envelope. A person can fail to qualify under a sound rule. A rule can fail to describe the person. Almost every institution I’ve worked inside produces identical paperwork for both, then treats the paperwork as the finding.3

The appeal leaves the category alone

An appeal works on the decision. It re-runs the determination against the same criteria, with the burden on you to show the criteria were misread. That matters, it’s the difference between a bureaucracy and a whim, but it leaves the category untouched by design, and the stamina it demands does most of the sorting before anyone decides anything at all.

Watch where the mismatch goes in the meantime. When a case contradicts the model, the cheapest move available is to reclassify the person instead of the model. Michigan’s MiDAS system issued unemployment fraud determinations without human review for two years. When the state finally looked at the files, the error rate ran past ninety percent, and in the meantime tens of thousands of people had wages garnished on the strength of a pattern match. Britain’s Post Office prosecuted hundreds of subpostmasters over shortfalls its Horizon accounting software had invented, because it read every discrepancy between a person and a computer as evidence about the person.

In both, following an internal process didn’t prove the process was sound. The applicant becomes the crumple zone, the part designed to absorb the impact so the machinery survives intact. Once the absorption works, the fraud story writes itself.

Who bends, and who can contest the rule

When a life collides with an administrative category, two things decide the outcome.

Who has to bend. Either the person bends to fit the rule, or the institution bends to fit the person.

And what you have standing to contest. Standing here means being able to force a system to treat your reality as an obligation, which is more than being believed. You might have standing over the decision. You might have standing over the rule that produced it. Different powers, and most systems grant the first while treating the second as out of scope.

That gives an institution four ways to respond.

Exclusion. The rule is fixed, nobody outside has standing, and the person absorbs the difference. During the Madras famine of 1876 the colonial administration kept grain exports moving under a doctrine that treated relief as market interference. Starvation without purchasing power registers as zero demand, so the ledger stayed clean while the population it described died.

Managed accommodation. The institution notices the mismatch and adjusts around it while keeping control of the terms. Predictive child-welfare screening does this: the model routes a household into monitoring, the agency may retune the model, and the family scored by it never gets any say over what the score is for.

Adversarial correction. You get standing over the decision and nothing more. Goldberg v. Kelly established in 1970 that welfare benefits could not be cut off without a hearing first. A real protection, won at cost. A claimant argues about fit; the statute that set the boxes isn’t a party to the argument.

Corrigibility. The mismatch counts as evidence against the category, and the people it excluded can force the category open. When ACT UP went after the FDA in the late 1980s, they weren’t asking for compassionate-use exceptions one patient at a time. They attacked the trial design itself, and the parallel track and accelerated approval pathways that came out of that changed what counted as evidence for everyone after.

Reform stalls between the third and fourth constantly.

Medicare shows what the stall looks like from inside one file. Beneficiaries and their advocates alleged that for years its contractors had denied skilled nursing and therapy claims because the patient was not improving, a rule of thumb found in no statute or regulation. Glenda Jimmo, a Vermont woman, is legally blind and lost part of a leg to diabetes. She had her home health care denied on that ground and became the lead plaintiff in a national class action. The settlement, approved in January 2013, had Medicare rewrite its manuals to say that a patient who needs skilled care to keep from declining is covered whether or not they can improve. The agency’s own fact sheet called this a clarification rather than an expansion of coverage, and noted that it had denied ever establishing an improvement standard.4 In April 2014 her own claim reached the Medicare Appeals Council, which upheld the denial because her condition was not improving. It took a second federal lawsuit to get the claim paid.5 In 2016 the court found that the government hadn’t accurately carried out the education campaign the settlement required. In February 2017 it approved a corrective statement disavowing the Improvement Standard by name.6 The rule had been sued, settled and written out of the manuals, and it was still deciding appeals more than a year later, because officially it had never existed.

What the PACT Act did instead

The standard response to a category that keeps hurting people is to build another lane of process around it: an ombudsman, a review board, a second-level appeal. Each lane taxes time and literacy and stamina, and the person least able to pay is the one who pays. Together they turn a defective rule into a private endurance test.

Compare how the same problem got solved for veterans. Proving that a specific cancer came from a specific burn pit is close to impossible for an individual, and for decades the VA let each dying claimant try. The PACT Act stopped requiring it: for more than twenty conditions, service in the relevant place and time now carries a presumption, and the institution takes on the causal question it had been offloading. Congress finally read thousands of individually unwinnable cases as one badly written rule.

An ordinary appeal asks whether the agency applied the rule correctly. Almost nobody has standing to ask whether the pattern of failures shows the rule is wrong.7

Building the threshold

Stability is the fair objection. If every losing claimant could reopen the statute, nothing would ever be settled. So the fix is a threshold rather than unlimited appeal: a point at which accumulated exceptions stop being individual variances and become an obligation to look at the category.

That requires the institution to keep the evidence in the first place. Near-misses and manual overrides and the cases a caseworker fixed by hand are currently treated as administrative residue, thrown out exactly where they’d be most useful. Logging them would let an agency see which criteria keep needing correction, and for whom.

The log also needs teeth. Clustered reversals on appeal have to trigger rulemaking instead of another round of one-off settlements, since paying individual claimants is how an agency currently buys the right to keep the rule. Assumptions and training data behind a scoring model have to survive long enough to be inspected. And consequential classifications need to expire on a schedule, so continuing to use one becomes a decision somebody makes instead of a default nobody revisits.

These give institutional commitments an enforceable form: a record, a trigger for review, and someone responsible for acting on it.8 Who may assemble the cluster, and what makes a finding bind once it exists, is the subject of Feedback Without Standing Fixes Only the Cases It Finds.

Diagram

Where an adverse decision can travel

View source

%% title: Where an adverse decision can travel
%% caption: Both paths start from a rule applied correctly. Only the threshold path can reach the rule itself.
flowchart TD

  classDef actor stroke-width:1.6px;
  classDef system stroke-width:1.2px;
  classDef gate stroke-width:1.6px;
  classDef binding stroke-width:3px,font-weight:bold;
  classDef status stroke-dasharray:4 3;
  classDef repair stroke-width:2.4px,font-weight:bold;
  classDef failure stroke-width:2.2px,stroke-dasharray:6 3;

  P([Person]):::actor --> A[Adverse decision]:::system
  A --> G{Appeal}:::gate

  G -- upheld --> D([Drop-off]):::failure
  G -- reversed --> R([Individual reversal]):::repair

  R -. closed as variance .-> V([Category untouched]):::status
  D --> E[Exception logged]:::system

  R ==> T{Cluster past threshold}:::gate
  E ==> T
  T ==> M[[Rulemaking obligation]]:::binding
  M ==> C([Category rewritten]):::repair

Which direction the information runs

The technologist’s version of this argument goes wrong in a specific way: if the institution’s picture of you is too thin, collect more of you.

The Dutch childcare-benefits scandal is the counterexample. The tax authority’s fraud model treated dual nationality as a risk signal and wrongly accused tens of thousands of families, driving parents into debt and children into foster care. That system was rich in data and structurally deaf. Transparency is not corrigibility: handing each family a printout of which coefficients ruined them would have changed nothing about their position.

Sometimes fidelity runs the other way and requires the state to know less. The secret ballot and the limits on biometric collection are friction deliberately preserved: they’re what keeps a person from being sorted before they have acted. Everything depends on which way it flows. Downward, the citizen becomes legible to the system. Upward, the system becomes answerable to the citizen. Only one of those is a democratic property, and collecting more will never produce it.

The polite version is the dangerous one

No government can abolish scarcity or tragedy. Budgets end, beds run out, majorities beat minorities. Political equality never promised you’d win.

It promises something narrower and harder: nobody gets to arrange your loss out of view to keep a diagram clean.

The agencies to worry about are often the courteous ones. They answer within one business day, log the complaint, confirm that the manual was followed, and carry on unchanged, because nothing a person actually lives through counts, in their process, as evidence against the rule that governed them.

A democracy has to categorize and decide; there’s no alternative. The open question is whether it can revise a category before repeated injuries force the issue. The PACT Act took decades and a lot of funerals. Michigan, Horizon, the Dutch families: still open.