In June 2020 the Social Security Administration's Office of the Inspector General published an audit of one step in the disability appeal path, and the agency it audits wrote back to say the numbers were wrong. The disagreement survives in the report, A-01-19-50762, the way institutional arguments usually do: politely, in an appendix, with both sets of figures printed. OIG estimated that of the 616,917 claimants denied disability benefits at the reconsideration level in calendar year 2015, about 86,400 — 14 percent — took no further action. SSA objected. Its own data, the agency wrote, "show that about 20 percent of individuals who filed their initial disability claims in CY 2015 did not appeal their reconsideration determinations to an ALJ." OIG printed the objection and answered it: "These figures are significantly different from OIG's estimates of 14 percent and 55 percent, respectively."

Neither side disputed that people who keep going are frequently allowed. OIG estimated that 290,000 of the 530,500 claimants who pressed on — 55 percent — later received allowance decisions; SSA's counter-figure was "around 47 percent received favorable decisions from an ALJ," lower because it counts judge decisions only. Two arms of one agency spent their argument on the size of the exit. That people win was never in question.

The appeal delivers for a large share of the people who complete it, and almost every feature of the path is built so that completing stays rare. This holds where federal law governs the process, as in marketplace health insurance and in disability. It holds where the law is newer and thinner, as on the platforms. The filter sits earlier than the point where anybody weighs the evidence. It sits at the moment a person decides whether the fight is worth a year of their life.

Start with the arithmetic, because the arithmetic is public. CMS requires insurers on HealthCare.gov to report claims and appeals, and KFF's analysis of the 2024 transparency files describes a funnel with an enormous mouth and almost no throat. Insurers received about 496 million claims in 2024, 451 million of them for in-network services, and denied roughly 85 million — a 19 percent in-network denial rate. Consumers appealed at least 262,982 of those denials, under 1 percent. Insurers upheld 165,863, or 66 percent of the ones they resolved, which leaves about a third of the denials that a human being bothered to contest reversed by the same company that had issued them.

Fix one thing in place before going further, because it decides an argument later on. That 85 million is already net of the easy cases. In CMS's accounting, a claim that was denied and then resubmitted and paid does not count as a denial at all. The number is not a census of everything that bounced. It is what was left after the corrections, the re-keyings and the second submissions had done their work.

The 2023 file, one year earlier under the same regime, rules out a fluke: 471 million claims received, 436 million in network, 86 million denied at a 20 percent rate, 376,508 appealed, 211,383 upheld. Fifty-six percent that year, sixty-six the next.

Walk down the 2024 cascade one order of magnitude at a time. Eighty-five million denied claims. Two hundred sixty-three thousand contested. A hundred sixty-six thousand of those contests lost. Then the last tier, which almost nobody reaches: a consumer whose internal appeal is upheld may have the right to an independent external review, decided by somebody who does not work for the insurer. In 2024, marketplace enrollees filed at least 5,881 external appeals, 4 percent of all upheld internal appeals. In 2023 it was about 5,000, or 3 percent.

The obvious reading of that last drop is that ninety-six percent of people who had already appealed once and lost then walked away — the cleanest evidence of attrition in the file, because these are proven fighters rather than a sample of the public. The obvious reading does not survive contact with the rule. Under 45 CFR 147.136, federal external review covers adverse determinations involving medical judgment — necessity, appropriateness, setting, level of care, effectiveness, the experimental and investigational — plus rescissions of coverage. Denials for eligibility, contractual exclusion, coding, or a missing prior authorization sit outside that scope. Some state processes reach further; the federal floor does not. And medical necessity accounts for about 5 percent of the denial reasons insurers report. So an unknown and probably small fraction of those 165,863 people had any external right to abandon, and the honest thing to say is that the abandonment rate at the last tier cannot be calculated from the public files at all.

It happens twice more.

The words "at least" are doing quiet work in all of those counts. CMS suppresses reported values under 10, and when KFF computes totals from the federal files it assumes a suppressed cell is zero, so every appeal figure above is a floor rather than a measurement. Then the same rule reaches the outcome that matters most. Reporting on external review in 2024, KFF writes: "Due to the suppression of small values, the rate at which external appeals were upheld could not be calculated." The final tier is the one where an independent expert can overrule your insurer. Its success rate cannot be computed from what CMS publishes, because so few people use it that the cells collapse under a privacy rounding rule. The design has driven a number so low that the number itself has gone missing.

Keep the denominators straight, because these systems do not share one. The insurance figures count claims, and a single sick person generates many. The Social Security figures count claimants, human beings with case numbers. An appeal rate under 1 percent of claims is not a statement that 99 percent of patients accepted their denials, and nobody should turn it into one. It is a statement about how many contested events, out of all contestable events, ever met a reviewer. Both denominators point the same way, which is what makes the conclusion robust to whichever one a reader prefers, not what makes it safe to add them together.

Why so few people use a remedy that works this well has one partial answer in KFF's 2023 consumer survey, which measures belief rather than entitlement. Asked whether they have a legal right to appeal to a government agency or an independent medical expert, 40 percent of consumers believed they did, 51 percent said they were unsure, and 9 percent believed they did not. Marketplace enrollees were least likely to know, at 34 percent, against 58 percent for people on Medicare and 45 percent on Medicaid. Among insured adults who reported a problem with their insurance in the past year, one in ten had filed a formal appeal. The most powerful remedy in the system is held by the population least likely to know it exists, and knowledge of it tracks what kind of insurance card you carry rather than what happened to your claim.

Disability filters differently. It filters with the calendar. A denied claimant must first ask the same state agency to reconsider, a second look that produces a second denial in the large majority of cases, and only then may the file reach an administrative law judge. Claimants who won at that reconsideration level in CY 2015 waited an average of 310 days for the allowance. Nearly a year, at the second of four levels, to be told yes.

Be fair to the tier, because the auditors were. Those 310 days were 535 days fewer than claimants waited in states where the first appeal went straight to a judge. Reconsideration allowed 84,866 people in CY 2015 and got them their answer far sooner than a hearing would have.

That comparison was available because of an experiment, and the experiment is the most interesting thing in the file. In 1999 SSA eliminated the reconsideration step in ten states — Alabama, Alaska, Colorado, Louisiana, Michigan, Missouri, New Hampshire, New York, Pennsylvania, and the Los Angeles North and West branches in California — so that a denied claimant went directly to a judge. It ran for nineteen years. Then, on 25 July 2018, SSA testified to a House subcommittee that it would put the step back. Its Deputy Commissioner for Analytics, Review and Oversight gave the reasons: "Reinstating reconsideration will restore uniformity to [SSA's] national programs. It will also provide claimants the opportunity to receive a favorable decision more quickly and will aid in alleviating the hearings backlog." Reinstatement began in January 2019 and finished in Alaska in March 2020. There are no prototype states now.

Those are real reasons, and the faster favorable decision is a real benefit. But read what SSA told its inspector general about the nineteen years: "Due to the changes from the original Prototype design over the years, we do not have any information documenting the results of Prototype." The agency ran a controlled comparison across ten states for nearly two decades, ended it, and does not hold the finding. The Social Security Advisory Board, asked by Congress to evaluate the same decision, reported that its efforts "were hampered by a lack of data or a decision not to provide available data to the Board."

The cost the audit did measure falls on the people the step does not help. For claimants who had to go on to a judge anyway, reconsideration added an average of 79 days: 924 days in a state carrying the tier, against 845 in a state without it. Two and a half years to a favorable decision from a judge, in a program whose applicants are by definition people who cannot work. The step returns time to the group it allows and takes time from the group it denies.

That is the friction stated in the only terms a household cares about. A person deciding whether to appeal is not being asked to assess the strength of their case. They are being asked whether they can survive another 924 days without income while they find out. The system does not need to reject them on the merits. It only needs to outlast them.

It works. Of the 616,917 people denied at reconsideration in CY 2015, OIG estimates 86,400 stopped entirely. Another 74,000 filed a new claim instead of appealing, which restarts the clock at level one. Of the 530,500 who went forward by either route, an estimated 290,000 received allowances. Count the new-claim filers as non-appellants, as an OIG footnote allows, and 26 percent of the denied population never appeared before a judge at all.

One caution the numbers demand. A later allowance does not establish that the earlier denial was wrong. A hearing is a fresh decision on a record that has changed — new medical evidence, the claimant's own testimony, a vocational expert, and, across 924 days, a condition that has often worsened. A judge can find disability with an onset date after the reconsideration denial, in which case the denial was right on what was in front of it. The defensible claim is narrower and still large: a very substantial share of people denied twice are subsequently allowed, and the argument here is not about who was right in 2015. It is about who never finds out.

The internal disagreement has to survive too. Take SSA's figures rather than OIG's — 20 percent walking away, 47 percent favorable at the judge — and nothing structural moves. One in five people denied at reconsideration abandons a process in which nearly half of those who reach a judge are allowed. The conservative pair sheds more people, not fewer.

Two systems, two filters. Insurance filters on knowledge: the remedy is fast enough on paper, and most people never learn the last tier exists. Disability filters on endurance: everyone denied receives a written path forward, and following it costs years that a person without income does not have. Neither filter operates on whether the original denial was any good.

The deadlines are where the asymmetry becomes physical, and they are specific rather than atmospheric. A marketplace denial must be appealed internally within 180 days. An external review must be requested within four months of the final internal decision, and the reviewer then has 45 days, and the outcome binds the insurer. In disability the window is 60 days at every level, running from receipt. There is an extension for good cause, which a claimant has to know exists in order to ask for it — the same shape as the external-review right that two-thirds of marketplace enrollees do not know they hold. A missed window converts a live dispute into a closed file, permanently, with nobody assessing whether the claim was good. That is a decision on the merits made by a calendar, and it is invisible in every dataset here, because a claim that dies on a deadline never becomes an appeal and therefore never becomes a row.

The economics point the same way, and they point there without anyone being a villain. A denial that gets reversed costs the organization the disputed amount plus the cost of handling the dispute. A denial that gets abandoned costs nothing, and in the accounts it looks identical to a denial that was right. No ledger in either industry separates the two. Every incentive that shapes the next version of the notice, the next staffing decision on the appeals line, and the next quarter's target is computed from a figure in which correctness and exhaustion are the same entry.

Even the advocacy is priced by the clock. An approved representative in a disability case is paid the lesser of 25 percent of past-due benefits or a capped amount, currently $9,200 for favorable decisions issued on or after 30 November 2024. Past-due benefits accumulate during the wait, so the system's private advocacy capacity is funded by its own slowness — and because the fee accrues on back pay, the economics reward taking a case to a hearing and barely reward taking one at reconsideration. That is the exact level at which the 86,400 disappear. Nobody designed that as an abandonment mechanism. It works as one.

So nobody has to intend attrition for the machine to run on it. They need only build a path with many exits and then decline to measure who leaves through them. An institution learns its abandonment rate the way a shop learns its shrinkage, as a stable background figure that budgets get built around. Once a denial rate is set on the assumption that under 1 percent will be contested, the assumption is load-bearing, and any serious effort to raise participation reads internally as a cost increase rather than an accuracy improvement.

Which brings us to the domain I expected to have no numbers at all.

Platform moderation is the case where an appeal is usually an appeal against a decision no person made — a classifier removes the video, and the review is not a second opinion but the first one. Until recently the honest thing to say was that the reversal rates were corporate property. That changed. The EU's Digital Services Act requires internal complaint-handling under Article 20 and, under Article 21, creates certified independent bodies to settle disputes out of court. Appeals Centre Europe, certified by the Irish media regulator, published its second transparency report on 28 May 2026, covering April 2025 to March 2026. It received more than 24,000 disputes, of which 12,400 were eligible, and issued more than 10,200 decisions, taking an average of 65 days.

Where it could see the content, it disagreed with the platform 59 percent of the time: 52 percent on removals, 63 percent on content the platform had left up after a report. On material reported as hate speech and left up, it overturned the platform in 70 percent of about 1,400 cases — TikTok 83 percent, Instagram 74, Facebook 61, YouTube 58. That is the insurance finding and the disability finding appearing a third time, in the place that was supposed to be unmeasurable: an independent reviewer, looking at decisions the issuing institution had already confirmed, disagreeing with the majority of them.

Now the part that belongs beside the CMS suppression rule. Of those 10,200 decisions, only about 2,900 involved reviewing the content. In more than 7,000 — 72 percent — the platform did not supply the material, and the Centre issued a default decision for the user without ever seeing what was in dispute. Account suspensions are the worst of it: more than 5,000 eligible disputes, and fewer than 150 decided with the content in hand. The BBC reported that Meta provided content in fewer than 100 of more than 4,600 account ban cases. The Centre notes that the share of cases where content arrives has "plateaued," that its decisions are not legally binding though platforms must engage in good faith, and that it knows of only a handful of cases where a platform has implemented one. For the rest, they "have either rejected our decisions or not responded at all."

So the third domain does have a statute, and a reviewer, and a published overturn rate. What it does not have is the number that would settle the question, because the party being reviewed declines to hand over the file.

Count them. In insurance, the external-review success rate cannot be computed, because a privacy rounding rule swallowed it, and the eligible population for that tier cannot be computed either. In disability, the agency that ran a nineteen-year experiment in ten states does not hold the results, and the board Congress asked to evaluate it could not get the data. On the platforms, 72 percent of independent decisions are made blind. Three institutions, three missing numbers, each absence sitting exactly where a finding would have gone, and each one owned by the party whose decision was under review.

Here is the strongest version of the objection, and it deserves the strong version. Low appeal rates are exactly what you would expect if most denials were correct. Insurance denials are largely administrative — a missing code, an excluded service, a referral nobody obtained — and the right response is a corrected resubmission from the provider's billing office, not a consumer appeal. There is a sharper form of it, too, and it is the one I find hardest: in an in-network denial the patient frequently has no direct financial stake at all. Network contracts and balance-billing rules leave the disputed money between the provider and the insurer. The rational actor is the billing office, not the person. A sub-1-percent consumer appeal rate might mean nothing more than that consumers are not the ones with an interest.

Part of that is answered by the way the number was built. Claims denied and then resubmitted and paid are not counted as denials in the first place, so the corrected code and the second submission are already gone before anyone prints 85 million. Every claim in that total is one a billing office either could not fix or did not try to fix. The objection describes a filter accurately and does not notice that the filter has already been applied. And where the professional does not act, the letter still arrives: the patients whose claims are quietly repaired never learn they were denied, and the ones without a billing office behind them get the notice and the choice.

What remains is answered by composition. CMS requires insurers to report reasons for in-network denials, and for 2024 they reported about 79 million of them. Thirteen percent were excluded services, 9 percent lacked prior authorization or a referral, and 5 percent were medical necessity. The two largest categories were "other," at 36 percent, and administrative, at 25 percent — more than half the reported reasons, between them, naming no clinical judgment at all. That is a description of the pool, not a measurement of error, and KFF is explicit that the share of claims denied for any given reason cannot be calculated from these files. But it dismantles the picture the objection needs. A pool in which contested clinical judgment accounts for one reason in twenty is not a pool of hard medical questions answered correctly by experts, and a person cannot rationally decline to contest reasoning they were never shown.

Then the selection argument fails on its own logic. Grant that appellants are unrepresentative — of course they are. The question is what they were selected on. If the filter sorted for merit, the appellant pool would hold the strongest cases and tell us little about anyone else. What the evidence describes is a filter that sorts for capacity. Awareness of external appeal rights varies by coverage type rather than by claim strength, 34 percent among marketplace enrollees against 58 percent on Medicare, and a right that half of consumers are unsure they hold cannot be declined on the merits by people who do not know they hold it. You cannot choose not to walk through a door nobody told you about. In disability the sorting variable is the ability to wait 924 days without earnings, a trait with no relationship to whether a person's spine is deteriorating.

Selection should still leave a fingerprint in the data, and it is worth going to look. It is not there, and the reason it is not there is instructive. In 2021 insurers upheld 59 percent of completed appeals. In 2023, with appeal volume four times higher, they upheld 56 percent. In 2024, with volume down by a third from 2023, they upheld 66 percent. Both columns move in both directions. These are also different populations by construction: each year's file covers only the plans still on sale two years later, 40 percent of the 2023 plans and 57 percent of the 2024 ones. There is a real hypothesis here — if drawing more people in lowered the average win rate, the unappealed pool would contain winnable cases nobody is bringing — and it is worth testing. It cannot be tested with what CMS publishes. That is a fourth missing number, and it is the one that would settle the central dispute of this essay.

The magnitudes are wrong for the mostly-correct story in one more place. Writing in Health Affairs, researchers examined independent medical review — the external tier — in the four states that publish its outcomes, across 2019 to 2023. Almost half of the coverage denials that reached it were overturned; a third of the cancer genetic testing denials that prompted the study were. Two caveats belong on that. It is a narrow window, and it is not limited to marketplace plans. And the tier hears mainly medical-judgment cases, the hardest and most genuinely arguable category, where a high overturn rate is less surprising than it first sounds. It is still the only sustained look anyone has published at what happens when a stranger reads the file, and roughly half the denials that get that far do not survive it.

Which turns the headline denial rate into something other than a measurement of what was owed. If a fifth of in-network claims are denied, and a third of the small minority that get contested are overturned, the honest reading is that the denial rate is calibrated to a population expected not to respond. That inference cannot be converted into a corrected rate, because nobody knows what the other 84.7 million contained. The uncertainty is an asset, and it belongs entirely to the party that issued the denial.

There should be a person in this essay. The brief asked for a reported, attributed account of somebody who appealed, and no such account was captured to the standard every figure here was held to, so none appears. The absence does structural work anyway. The records that carry this case are records of dispositions: claims received, claims denied, appeals filed, appeals upheld, days elapsed, allowances issued. The 86,400 people who stopped after their reconsideration denial exist in the public record as a sampled estimate in the appendix of an audit their own agency disputed. They left no hearing transcripts and no case numbers that lead anywhere. The winners generate files. The people who gave up generate a rounding difference between two agencies — and the rounding difference is the whole point. OIG's 14 percent is about 86,000 people. SSA's 20 percent is about 123,000. The distance between two offices of one agency is some 37,000 human beings, in one year, at one step of one program, and not one of them was ever asked why they stopped.

It would also be the easiest thing in the world to invent one. A composite claimant with a plausible town, a plausible diagnosis, and a heartbreaking detail about the third letter would read exactly like a real one, and no reader could tell the difference. That interchangeability is worth sitting with, because the machine described here runs on a related trick: a notice that is technically complete, formally accurate, and functionally inert. The forgery and the disclosure work the same way. Both satisfy the reader checking whether the box was ticked. Both fail the reader who needed the information.

Return to where all of this starts, which is a letter. The notice that denies your claim is usually telling the truth about your rights, in the same paragraph as the bad news, and the truth is favorable to you. About a third of contested marketplace denials came back overturned in 2024, and 44 percent the year before. Where an independent body in Europe could actually see the material, it disagreed with the platform in 59 percent of cases. Nearly half the people who reach a disability judge are allowed. None of it is hidden. It is printed in the government's own files and argued over by the government's own auditors.

What is missing is never the rights. It is always the count. The letter tells you the truth and then the institution declines to learn what happens next — how many people read it and stopped, what share of them were right, how the nineteen-year experiment came out. A process can be accurate at every step and still be built so that most of the people entitled to use it do not, and the surest way to keep that arrangement is to leave the one number that would expose it uncollected. No conspiracy is required. It is a filing decision, made the same way a thousand times, in a room where nobody has to decide anything at all.