1. A reversal deserves an explanation before it becomes a rate
An appeal succeeds and a post comes back. That is a meaningful event for the person who posted it. It may also be evidence of a mistaken removal. But a public statistic counting successful appeals cannot, by itself, establish how often the original system was wrong. The people who appeal are a selected group, some cases remain unresolved, and the reason for a changed outcome may be more complicated than a reviewer reversing a mistaken interpretation of an unchanged rule.
The scale is substantial. On February 17, 2026, the European Commission reported that 30 percent of 165 million moderation decisions challenged through platforms' internal mechanisms had been reversed. It separately reported that out-of-court bodies overturned platform decisions in 52 percent of closed cases in a first-half-2025 collection involving Facebook, Instagram and TikTok. These are attributed Commission figures with different populations and procedural stages, not two interchangeable estimates of overall platform error. The Commission's February 2026 announcement
A serious account should preserve what those figures make visible: people challenged decisions and some outcomes changed. It should then ask what the counting system can establish about the original decision, the appeal process, and the remedy delivered. Treating every reversal as proof of misconduct is too strong. Treating reversals as an unimportant detail is too weak. A fair system needs both a way to correct mistakes and a way to learn what produced them.
This article develops an original fictional service called Forum R. Its examples are not estimates of a real platform's performance. No private moderation records were obtained, no real appeal was submitted, and no model or human moderation team was tested. One small local script checks an invented case ledger and deliberately introduced data errors. The exercise is a test of that ledger, not an audit of a company's production system.
The focus is the appeal pathway, rather than the separate question of how platforms count removed posts or harmful-content exposure. We will follow cases from an initial action through submission, review, outcome, and observable restoration. Along the way, the same apparent success rate will conceal different patterns of access and delay. The aim is to write a conclusion that tells a user what changed and tells an investigator what remains unknown.
The central distinction is between three claims: the outcome changed, the earlier decision was wrong under a specified standard, and the user received an adequate remedy. A single case can support all three, only one, or different answers at different times. Keeping those claims separate makes accountability more demanding, not less. It prevents a favorable outcome label from doing work that only the underlying case record can perform.
2. Start with the kind of action being appealed
Reddit's report for July through December 2025 distinguishes appeals of content-level sanctions from appeals of account-level sanctions, reporting 481,416 and 499,515 respectively. It describes granted appeals as resulting in reinstatement of the account or content. These are the company's reported figures and categories for that period, not an independent assessment of each decision. The distinction is useful before making any calculation: a content item and an account restriction are different objects. Reddit's second-half-2025 transparency report
Forum R begins with a fictional account that contains 40 posts. An account suspension makes all 40 inaccessible. A successful account appeal restores access to the account and those posts. A report could legitimately count one resolved account appeal and 40 posts becoming available, provided it names both measures. Calling the episode 40 successful appeals would invent 39 complaints. Calling it a single restored post would conceal most of the practical effect.
Now suppose three of the 40 posts also have separate content restrictions. Restoring the account does not logically settle those separate actions. In this constructed example, 37 posts become accessible and three remain restricted. A simple account-status check therefore cannot establish that everything returned. The case record needs the appealed action's identifier and the objects affected by that action, not only the account name and the word restored.
The reverse problem occurs when one post has two independent restrictions. Forum R removes a fictional image under one policy and separately limits the account under another. The image appeal succeeds, but the account restriction still prevents public access. The content decision changed while the user-visible state did not. That outcome may require a second explanation, but it is not necessarily evidence that the first reviewer failed to implement the decision. The investigator must locate the remaining constraint.
A useful appeal record therefore begins with an action, not a person's general dissatisfaction. It should identify the disputed measure, its target, the scope of the requested remedy, and the other restrictions known to affect that target. This is an analytical proposal for the fictional exercise, not a claim that every service uses the same schema. Its purpose is to prevent an account event from being mistaken for a content event or one restriction from being mistaken for all restrictions.
The action-level approach also avoids inflated person counts. One person can appeal several decisions; several people can be affected by one shared account's restriction. A count of appeal records is not automatically a count of distinct individuals denied a voice. The practical consequences deserve attention, but they must be measured or described with evidence rather than inferred from a convenient identifier.
3. Preserve more than yes and no
YouTube's current help page distinguishes warnings or strikes, which can be appealed for six months, from content removals, for which it states a one-year appeal window. It also describes outcomes other than unrestricted reinstatement, including age restriction, and distinguishes this process from copyright appeals. These details were checked on the current page; they are not a promise that every type of restriction follows one universal appeal procedure. YouTube's Community Guidelines appeal guidance
Forum R's first invented cohort contains 1,000 submitted appeals. At the observation date, 800 are resolved and 200 are pending. Of the resolved cases, 240 change the original outcome and 560 leave it unchanged. Dividing 240 by 800 gives a 30 percent change rate among resolved cases. Dividing 240 by all 1,000 submissions gives 24 percent with a changed outcome so far. Those are different questions, and neither should quietly replace the other.
The 240 changed outcomes are then divided according to the fictional review notes. In 120, the reviewer identifies a mistaken application of the same rule to the evidence already available. In 60, new contextual information changes the assessment. In 40, a policy revision changes what is permitted. In 20, a discretionary accommodation changes the action without resolving the original classification. The categories sum to 240. They were assigned for this example; they do not purport to describe the distribution at any real service.
Calling all 240 cases original classification errors would erase those distinctions. Calling only the first 120 important would miss the value of new context and the consequences of policy choices. A useful report can show both the broad outcome-change count and the narrower reason categories. It must also say who assigned the reasons and whether those assignments were independently checked. A reason code is evidence of a recorded explanation, not automatic proof that the explanation is correct.
Partial outcomes need their own state. Suppose a separate fictional case changes a complete removal to a warning screen. A binary export might call that successful because the original action changed, or unsuccessful because access remains limited. Neither label alone tells the reader what happened. The record should preserve the before-and-after states and let the report define any summary category explicitly.
There is no need to force every case into a grand theory of censorship or safety. For a particular user, the concrete question may be whether a video is accessible, whether an account strike remains, or whether a restriction was narrowed. For an evaluator, the question may concern how an unchanged policy was applied. Those purposes overlap, but they should not be collapsed into a single outcome flag that becomes increasingly vague as it travels into a headline.
4. Count restoration separately from appeal decisions
In a May 2019 explanation of its reporting, Facebook described restoring content both after appeals and without a direct appeal. Its examples included extending a correction to other copies of the same content and identifying an error before the user appealed. This is a historical first-party description, not a fresh claim about today's totals or workflow. It establishes why a restoration counter need not equal a successful-appeal counter. Facebook's May 2019 reporting explanation
Keep Forum R's 240 changed appeal outcomes, but now stipulate that each concerns one directly appealed post and results in that post being restored. A separate proactive review restores 110 additional posts that had no individual appeal. The service therefore restores 350 posts while recording 240 changed appeal outcomes. Dividing all 350 restorations by 800 resolved appeals produces 43.75 percent, but that ratio is not the appeal-change rate. Its numerator includes events outside the appeal cohort.
One correction can also affect several objects, as the account example showed. To reconcile the numbers, Forum R uses an action identifier, an appeal identifier, and a content identifier rather than treating them as the same field. A linkage record says which appeal, if any, triggered each restoration. The comparison becomes checkable: there are 240 directly appealed restored posts and 110 restored through the separate review route in this constructed scenario.
The practical state after restoration still needs observation. In another stipulated case, the internal decision says restore at noon, the enforcement restriction is cleared at 12:05, and public access is confirmed at 12:40. These are three different events with a 40-minute interval from decision to observed access. An internal success message at noon cannot establish that users could already see the post. Nor can a public check at 12:40 establish the exact second when availability first returned.
Forum R's reader-facing report records the observation method. A logged-out availability check answers a different question from a creator seeing the post in a management interface. The fictional example does not actually perform either check on a real service. It specifies the distinction because a researcher who claims restoration should explain what was observed, with which account state and at what time. Otherwise an administrative field may be mistaken for a complete user experience.
This separation creates a useful diagnostic. If the decision record changed but the post remained unavailable, the next inquiry concerns implementation or another restriction. If the post returned but the appeal record did not change, the next inquiry concerns linkage or the route by which it returned. Neither pattern supplies a motive by itself. Both patterns identify a narrower system behavior that can be investigated without turning a mismatch between counters into an unsupported allegation.
5. The people who appeal are not a random sample
Imagine Forum R issued 10,000 removals under one fixed rule. Exactly 1,000 affected decision records entered the appeal cohort described above. Even if the reviewer classified every appealed case perfectly, the results would not automatically reveal the error rate among the remaining 9,000. The appeal process has selected cases according to user behavior and access, not according to a probability sample designed to estimate population accuracy.
Two invented worlds make the limit explicit. In World One, the 120 original-rule mistakes identified through appeal are the only such mistakes among all 10,000 removals. The population mistake fraction would be 1.2 percent. In World Two, there are another 1,800 original-rule mistakes among the 9,000 unappealed removals. The total becomes 1,920, or 19.2 percent. Both worlds preserve every fact in the observed appeal ledger. The ledger alone cannot distinguish them.
These are logical possibilities, not estimates or a statistical confidence interval. They do not imply that either world is likely. Their purpose is to show that an inference needs information about the unappealed population. Even the first world's interpretation assumes the review labels are valid and the meaning of original-rule mistake is fixed. If those assumptions are uncertain, the population conclusion becomes less secure rather than more.
Appeal selection can operate in more than one direction. A person with a clearly permitted post may abandon the process because the post is no longer useful. Someone who deliberately violated a rule may appeal because there is little cost to trying. A person may not understand the notice, may have lost account access, or may disagree with the rule rather than the facts. These are possible mechanisms, not measured explanations for any named platform. The observed appeal rate cannot tell the analyst their relative importance.
A separate sample review could address the population question if its design actually covers both appealed and unappealed decisions. Forum R's proposed evaluation would select decisions from a defined initial-action population, retain selection probabilities, and evaluate them under an explicitly chosen policy version. It would not simply add more appealed cases and call the larger collection representative. The size of a selected dataset does not remove the selection process that created it.
This distinction is especially important when a platform or critic claims improvement. A falling reversal rate could reflect more accurate original actions. It could also reflect a change in who reaches review, the kinds of cases selected, or the evidence available at appeal. The proper next step is to compare those mechanisms with evidence. The rate is a signal to investigate, not a self-interpreting verdict about the original moderation system.
6. A fast median can coexist with a slow unresolved queue
Forum R's 800 resolved appeals have a deliberately simple timing distribution. Six hundred finish after one day and 200 finish after nine days. The median among resolved cases is one day. Meanwhile, all 200 still-pending appeals have been waiting for 25 days at the observation date. Reporting the one-day median without the pending queue would describe completed work accurately while giving an incomplete impression of the experience of everyone who submitted an appeal.
The unresolved cases do not yet have final processing times. Assigning them 25 days as if that were their total duration would understate any additional waiting. Excluding them from a completed-case median is mathematically normal, but the report should then disclose their number and current ages. An analyst should not fill an unknown completion date with the observation date and treat the resulting duration as a completed outcome.
The service also receives cases from earlier periods. In the same reporting window, it closes 200 older appeals, 160 of which change the outcome. Across all work completed during the period, it closes 1,000 cases and changes 400 outcomes: 240 from the new cohort and 160 from the older cohort. That is a 40 percent change rate for period completions, compared with 30 percent among resolved new-cohort cases. Both figures are correct under their definitions.
If a report pairs the 400 period changes with only the 1,000 newly submitted cases, it may appear to produce a meaningful 40 percent new-cohort result, but it has mixed memberships. The arithmetic coincidence of equal totals conceals the mismatch. The older cases belong in a workload account for the period, not in the outcome history of the new submissions. A case-level join, rather than a comparison of aggregate totals, is needed to detect the error.
Delay also has a consequence that a duration alone cannot summarize. Suppose an invented notice about a community meeting has 18 hours of usefulness remaining when it is removed. It returns 30 hours later. The appeal may have corrected the action, but the notice missed its original practical window by 12 hours. This does not make every 30-hour review unacceptable; it shows why timeliness and eventual outcome can answer different questions about a remedy.
The report should therefore make room for current queue age, completed processing time, and the time-sensitive purpose of affected content where that purpose is actually established. It should not infer harm from a timestamp alone, nor erase a documented loss of usefulness because the final status says restored. An appeal process is experienced over time. Its evaluation should preserve that timeline rather than compress it into one favorable endpoint.
7. Access can change the number of corrected cases without changing the review rate
Consider two hypothetical versions of Forum R's appeal interface. Each serves 1,000 eligible action records. In Version A, appeal information is viewed for 800 action records, 200 records proceed to submitted complaints, and 40 receive changed outcomes. In Version B, information is viewed for 400 action records, 100 records proceed to submitted complaints, and 20 receive changed outcomes. Assume for this exercise that all submitted cases are resolved. Both versions report a 20 percent change rate among appeals, but Version B produces half as many changed outcomes.
This is an illustration, not an actual interface experiment. It does not establish that visibility of information caused the difference. The two populations could differ in other ways. To identify a causal effect, an evaluator would need a suitable study design and safeguards, not merely two convenient totals. What the example proves arithmetically is that an unchanged review-outcome percentage can coexist with a substantial difference in the number of action records reaching the process and obtaining a changed outcome.
The intermediate counts explain where further investigation belongs. If the notice was delivered but its link did not open, the issue is not the same as a user deciding not to appeal after reading a clear explanation. If the form opened but submission failed, the issue differs from a complete submission waiting for review. Forum R's proposed evaluation records these stages separately so a low submission count is not automatically interpreted as satisfaction with the original decision.
Care is needed with the word reachable. In the hypothetical study, it means a specified observable event, such as successfully loading the relevant information page. It does not mean that the user understood the explanation, could complete the form, or believed an appeal would help. Those would require separate evidence. An event log can establish that a page loaded while leaving the experience and comprehension of the person unknown.
A privacy-conscious account can investigate these stages without publishing identifiable case histories. The exercise can use aggregate stage counts and, where individual examples are necessary, authorized and appropriately protected records. Researchers should not solicit private account credentials or expose a user's restricted content simply to demonstrate that an interface was confusing. The public-interest question is how the mechanism works, not how much private material an investigator can gather along the way.
For an affected user, a clear explanation and working submission path matter even when the ultimate decision remains unchanged. For the platform, reduced friction may initially increase complaints and reversals by revealing cases that previously never reached review. That increase would not automatically mean initial moderation had deteriorated. A reporting system that interprets every increase in appeals as failure could create an incentive to make the process less usable, precisely the opposite of what an accessible correction mechanism should accomplish.
8. Human involvement needs evidence about the review itself
The second iteration of the Santa Clara Principles, developed in 2020 and 2021, recommends accessible appeals, contextual competence, reasoned outcomes, and human review by people not involved in the initial decision. These are civil-society recommendations, not a universal legal rule or proof that a company's current process meets them. They help frame what a meaningful review might need beyond a button labeled appeal. The Santa Clara Principles' Appeal section
A human-reviewed label describes participation by a person, but it does not independently establish the correctness or depth of the review. In Forum R's imagined records, one reviewer sees the original content and the user's explanation. Another sees only a machine-generated summary that omits the disputed context. Both might be counted as human involvement, while the information available for the decision differs materially. The account should not infer equal review conditions from an equal label.
A concrete quality check could focus on one narrow class of omission. Construct 12 fictional case packets, each with an initial rationale and a separate appeal statement. Six statements contain a specified contextual fact that changes the case under the exercise's stipulated rule; six do not. A test of the review interface could check whether all 12 statements are displayed completely and attached to the correct case. That would test information delivery, not whether actual moderators reached fair judgments. No such interface test was performed for this article.
Suppose the hypothetical delivery check finds that two statements were truncated. The first necessary conclusion is that two packets were incomplete under the test. It is not that two users necessarily deserved reversal. The investigator would then need to determine whether the missing text mattered to the relevant rule and whether it was available through another part of the workflow. This sequence prevents a technical defect from being inflated into an unsupported outcome count while still treating the defect seriously.
Review consistency is another distinct question. If two independent reviewers examine the same invented set of ten cases and disagree on three, the observed agreement is seven of ten, or 70 percent. Neither reviewer becomes the truth standard merely because their decision was recorded first. A reconciliation process would need to examine the reasons for disagreement, the clarity of the rule, and the adequacy of the evidence. A majority vote can settle an administrative outcome without proving that the underlying interpretation is uniquely correct.
These distinctions matter when comparing automation with human review. A low appeal-change rate after human handling could reflect reliable initial decisions, a strict review standard, limited evidence, or shared errors. The human label does not choose among those explanations. A useful evaluation identifies what information reviewers received, which standard they applied, and how their judgments were checked. The relevant question is not whether a person appeared somewhere in the workflow, but whether the additional review supplied a meaningful opportunity to reconsider the disputed issue.
9. Selected external cases can reveal a problem without measuring its prevalence
The Oversight Board's 2024 annual-report highlights, published August 27, 2025, reported 65 decisions and 558,235 appeals, including eight cases referred by Meta. The decisions included standard, expedited, and summary forms. These historical counts describe different stages and units; they should not be converted into a simple probability that an individual appeal received a full review. Nor do the highlighted outcomes alone measure the error rate across all content on Meta's platforms. The Board's 2024 annual-report highlights
An external case can be important because it illuminates a difficult rule, a recurring issue, or a consequential failure. That value does not require the case to be statistically representative. The mistake arises when an analyst turns the outcome of a selected case collection into a population estimate without establishing how cases entered the collection. A powerful example and a probability sample serve different purposes.
Imagine Forum R sends ten particularly ambiguous cases to a fictional external panel. The panel changes eight outcomes. Separately, a defined random sample of 100 initial decisions receives independent review under the exercise's fixed standard, and 20 are classified as incorrect. The first result is eight changes among ten selected cases. The second is 20 classifications among 100 sampled decisions. Comparing 80 percent with 20 percent as if they measured the same population would erase the selection rules and the different outcome definitions.
The selected panel might still uncover an important recurring defect. Suppose five of its cases turn on the same missing translation context. That provides a reason to investigate the translation pathway. It does not establish that half of all moderation decisions suffer from the defect. A follow-on evaluation would need to define the population exposed to the pathway and inspect an appropriate sample or complete relevant inventory. The case finding generates a testable question; it does not supply the answer to that broader test.
A policy recommendation is also a different output from an individual reversal. Forum R's panel could uphold a particular action while recommending clearer notices for similar cases. Alternatively, it could reverse one action without recommending a general policy change. A dashboard that counts only reversals would miss the first type of contribution. A dashboard that counts recommendations without tracking implementation would miss whether anything changed afterward.
The public account should therefore follow the chain from selected case to identified issue to proposed change to observed implementation, stopping where the evidence stops. It should not assume that issuing a recommendation fixes a problem or that the absence of a recommendation validates every similar decision. Selected review can expose a mechanism worth examining in depth. Its contribution is strongest when that mechanism is described precisely enough for someone else to test beyond the original case.
10. Keep the reporting version attached to the number
YouTube's reporting change log records a May 2024 correction to video reinstatement data for July through September 2023, from 31,183 to 26,483. The same log records later changes, including September 2026 updates to presentation and category treatment. The older correction is a documented historical reporting revision, not evidence of today's error rate or an allegation about intent. It shows why an appeal-related number needs a reporting version as well as an event period. YouTube's transparency-report change log
The difference between those historical reinstatement figures is 4,700. Relative to the original 31,183, the reduction is approximately 15.1 percent. That calculation describes the size of the published correction. It does not mean 15.1 percent of moderation decisions were wrong, that 4,700 users lost appeals, or that the revised figure has been independently reconstructed here. A change to a reporting total must not be transformed into a claim about an unobserved operational event.
Forum R's own fictional export demonstrates a narrower versioning problem. Its first reporting definition counts only complete restoration as a changed outcome. A later definition counts both complete restoration and a reduction in restriction. Under an invented set of 100 resolved cases, 20 are fully restored, ten receive a reduced restriction, and 70 remain unchanged. The first definition produces 20 percent; the second produces 30 percent. The underlying case outcomes have not changed.
To compare periods, the analyst could recalculate both under one definition if the necessary case states are preserved. If only aggregate totals survive, that reconstruction may be impossible. The report should then mark a definition break rather than invent a conversion factor. Knowing that partial outcomes exist is not enough to determine how many occurred in an earlier period where they were not separately counted.
Versioning also applies to the policy used to judge the initial action. Suppose a fictional post was prohibited under Rule One but allowed under Rule Two after a documented policy change. A later restoration may be a correct implementation of Rule Two without proving that the earlier action violated Rule One. An investigator can still criticize Rule One on substantive grounds, but that is a different claim from finding an inconsistent application of the rule then in force.
The most useful appeal record preserves the policy version, evidence available at each stage, and reason for the changed outcome where the record supplies one. That permits separate questions: did the initial reviewer apply the rule consistently, did new evidence alter the assessment, and did the rule itself change? A reporting system that retains only the final status cannot reliably answer all three. A well-documented revision improves the record; an unexplained change should remain a visible question rather than an invitation to invent its cause.
11. Run a small ledger through checks that can actually fail
To make the accounting concrete, this article uses 12 invented cases, C1 through C12. Eight have a merits decision, two are pending, one was withdrawn, and one is outside the exercise's eligible process. Among the eight decided cases, five change the original action and three uphold it. Of the five changes, three restore full access, one reduces the restriction, and one has not yet produced restored access because a separate account restriction remains. These assignments are fictional, not extracted from a platform.
A local Python script actually checks this ledger. It counts the eight merits decisions, five changes, and three full-access restorations, producing a 62.5 percent change rate among decided cases. Five changed outcomes divided by all 12 submitted records gives approximately 41.7 percent changed so far, but that denominator includes pending and non-merits dispositions. Three full restorations divided by five changed outcomes gives 60 percent in this example. Each result is labeled according to its own question.
The first deliberate control duplicates C1, a changed case. A row counter now sees 13 records, nine merits decisions, and six changed outcomes. Its apparent change rate becomes approximately 66.7 percent. A uniqueness check on the case identifier rejects the duplicate instead of allowing an extra row to become an extra appeal. This is an observed result of a custom local script, not proof that a real platform's exports contain the same defect.
The second control changes C7's outcome to changed while leaving its decision date absent. The checker flags the inconsistent state because this exercise requires a decision date for every merits outcome. That rule is specific to the constructed schema. A different system might represent provisional decisions differently. Passing or failing the checker establishes compliance with the stated local rules, not the quality of a real moderation decision or conformance with a law.
The base fixture also keeps restoration state separate from outcome. C11 has a changed decision but no full-access restoration, while C4 has only a reduced restriction. A script that automatically maps every changed outcome to restored would generate five full restorations where the fixture contains three. This is the kind of error a set of plausible totals can conceal unless the relationship between fields is explicitly tested.
The exercise's limits are as important as its successful checks. It does not determine whether C1's assigned original-error label is justified, because the fictional ledger contains no real content for independent evaluation. It does not test a website, submit a complaint, or measure reviewer behavior. It tests whether a small authored record set preserves distinctions that the article relies on. That modest test is useful precisely because its claim is narrow enough to reproduce and because its controls show that the checker can detect the particular defects it was designed to detect.
12. Report both correction and the limits of correction
A defensible summary of Forum R would not say simply that 30 percent of removals were wrong. It would identify the initial action population, the submitted appeal cohort, the resolved cases, and the outcome definition. It would show the pending queue separately and distinguish a changed decision from a completed user-facing remedy. The result would be more detailed than one headline number, but each detail would answer a question that materially changes the interpretation.
For the 1,000-case cohort, the account would report 800 resolved appeals, 240 changed outcomes, and 200 still pending. It would preserve the assigned reasons for change rather than rename all changes original errors. It would identify the separate 110 proactive restorations rather than fold them into the appeal numerator. It would state that no representative review of the 9,000 unappealed removals had been performed, so the overall initial-error rate remained unknown.
The report could also say something concrete about process. Completed cases had a one-day median under the invented distribution, while the pending cases had already waited 25 days. That combination identifies a queue worth investigating without asserting why it occurred. The next question might concern prioritization, case complexity, missing evidence, or capacity. The record should not select one cause merely because it is politically convenient or intuitively plausible.
For users, a successful remedy may require more than a favorable review message. A post may return after its useful moment, an account-level restriction may remain, or a partial reduction may leave a significant burden. Those possibilities do not make appeals pointless. They define the additional outcomes an evaluation should examine when evidence is available. Correction is valuable even when incomplete, but incompleteness should not be hidden by a success label.
For institutions, an increase in reversals can sometimes mean that a correction pathway is reaching cases it previously missed. A decrease can sometimes mean that initial decisions improved. Neither interpretation follows automatically from the direction of the graph. The investigator's task is to connect the movement to evidence about the relevant mechanism, while preserving alternative explanations that the data do not eliminate.
The most useful public question is therefore not whether appeals prove that platforms are always wrong or usually right. It is whether the available record lets us distinguish initial decision quality, meaningful reconsideration, and delivered remedy. Those distinctions protect users from being reduced to a favorable percentage and protect the analysis from claims its evidence cannot support. An appeal is part of the system being evaluated. It should be measured as a process with access, decisions, delays, and consequences, not treated as a convenient correction factor attached to the original action count.