1. An approval and an answered question are not identical records
A drug receives accelerated approval. Years later, a study finishes, a paper appears, and a database entry changes. Which event establishes that the anticipated benefit was verified? The answer cannot be read from whichever date is easiest to find. An approval, a completed trial, a submitted report, a published result, and a regulatory determination are different events. Following the evidence requires connecting them without treating any one as a substitute for all the others.
FDA describes accelerated approval as a route for certain serious conditions using a surrogate or intermediate clinical endpoint reasonably likely to predict benefit, with further studies to confirm the anticipated benefit. The pathway allows earlier access while leaving a defined evidentiary question to be addressed. It is not a statement that every outstanding question has already been resolved, nor a declaration that the product has no benefit. FDA's patient explanation of accelerated approval
The public-interest question is what happened after that initial decision. Did the required work start? Did it address the intended population and outcome? What did the results show, and what did FDA subsequently decide? A good account can distinguish delay from completion, evidence from a status label, and a change affecting one use from a change affecting an entire product. A poor account flattens them into a binary story about a drug being either proven or disproven.
This article is a guide to regulatory records, not medical advice or a recommendation to start, stop, or switch treatment. Individual care requires a qualified clinician who knows the patient's circumstances and current evidence. No real patient's records are used here, and no named medicine is assigned a new efficacy or safety verdict. The examples use fictional products, indications, trials, and documents so that the reasoning can be inspected without pretending to conduct a clinical review of a real therapy.
The running example is a small invented program called Cedar. It contains eight products, 12 accelerated-approval indications, 15 postmarketing requirements, and ten trial identifiers. Those numbers are deliberately different. A local script checks the relationships and selected failure cases. It does not evaluate a medicine, test FDA's databases, or determine legal compliance. It asks whether a proposed public account keeps its objects and evidence aligned.
The payoff is a more precise final sentence. Instead of saying that a study was published and therefore the drug was confirmed, an account should identify the indication, the relevant requirement, the study that addressed it, and the regulatory action supported by the record. If one of those links remains uncertain, the uncertainty belongs in the finding. That does not prevent scrutiny. It locates the exact question scrutiny needs to answer.
2. Follow the indication rather than only the brand name
FDA's ongoing oncology accelerated-approval table is organized around indications and their associated postmarketing requirements. It directs readers to the latest approvals and prescribing information, and warns that an indication can remain on the ongoing page until labeling is updated or a Federal Register notice records a status change. The table is therefore a useful index, not a guarantee that every displayed row captures the latest regulatory event at the instant a reader visits. FDA's ongoing cancer accelerated-approval table
Cedar's fictional Product P1 has two indications. I1 is a particular use in Population A after prior treatment; I2 concerns a different use in Population B. The packet assigns verified benefit to I1 and an ongoing confirmation process to I2. A product-level dashboard that labels P1 verified would hide the unresolved indication. A dashboard that labels it unverified would hide the completed pathway for I1. The problem is not which single word to choose; it is that the proposed unit is too coarse.
Another product, P2, has one verified indication and one withdrawn indication in the invented ledger. Reporting that P2 was withdrawn would overstate the scope. Reporting only that P2 remains approved would conceal the withdrawn use. The accurate account names the indication and the action that concerns it. This remains true even when readers are more familiar with a brand than with the detailed population and treatment setting in the regulatory record.
The indication should be represented by enough information to prevent an accidental match. For the exercise, that means product identifier, population, disease setting, prior-treatment condition, combination regimen if applicable, and the particular approval event. These fields are an original bookkeeping proposal, not a complete legal definition of an indication. Their purpose is to make visible the differences a short brand-name label would discard.
A trial title alone is also insufficient. Suppose Cedar's T2 study concerns Population B before any prior treatment, whereas I2 concerns patients after prior treatment. The similar product and disease name do not establish that T2 satisfies I2's requirement. There may be a documented rationale connecting the populations, or the study may concern a different regulatory question. The analyst needs the actual requirement and the agency's account of how the evidence addresses it before making the connection.
This is why a careful investigation starts with the approval letter and indication text, not a search for the product followed by the first positive paper. The latter approach can produce a true statement about the wrong use. The former preserves the question the approval pathway actually left open and makes later evidence easier to evaluate against that question.
3. Reconcile products, requirements and trials before counting outcomes
Cedar's 12 indications belong to eight products. P1 and P2 each have two indications; P3, P4 and P5 each have one; P6 has two; P7 has one; and P8 has two. The counts add to 12. Five indications are assigned verified status, two withdrawn status, and five ongoing status. These are invented classifications chosen to demonstrate an accounting problem, not a survey of real approvals or an estimate of the pathway's success.
The product-level sets overlap. P1 through P5 each have at least one verified indication, giving five products in that set. P2 and P6 each have a withdrawn indication, giving two in that set. P1, P6, P7 and P8 each have an ongoing indication, giving four in that set. Adding five, two and four gives 11 category memberships, not 11 products. There are only eight distinct products. Treating the memberships as mutually exclusive would create three products that do not exist.
The requirements add another layer. In this fictional ledger, each indication has at least one requirement. I1, I9 and I12 each have a second requirement, producing 15 requirement records. The records map to ten distinct trials because some requirements concern different reports or follow-up components of the same trial, and some trials connect to more than one indication. A count of requirements is therefore not automatically a count of independently conducted experiments.
For example, I1's two requirements both refer to T1 in the constructed packet. One concerns the principal report, while the other concerns a later follow-up report. Marking the first requirement complete cannot automatically close the second. Conversely, counting both as two separate trials would overstate the number of independent study populations. The relation must be preserved even when the report ultimately summarizes the findings in ordinary prose rather than a technical database.
The local exercise actually checks these identifiers and relationships. It confirms eight product identifiers, 12 indication identifiers, 15 requirement identifiers, and ten trial identifiers under the authored mapping. It also tests a duplicate requirement row and an unknown indication reference. The first must not become a sixteenth unique requirement; the second must not silently create a new indication. These are bookkeeping checks, not judgments about whether any requirement was scientifically adequate.
A public conclusion should keep the chosen unit visible. Five verified indications out of 12 says something different from five products with at least one verified indication out of eight. Neither is a patient-level probability of benefit, and neither summarizes the quality of the underlying evidence. The reconciliation simply prevents the article from asking one question and answering it with a count of a different object.
4. Check what the endpoint can establish in its particular setting
FDA's December 2018 oncology endpoint guidance distinguishes survival, tumor-response measures, progression-related measures, and symptom assessments. It defines overall survival from randomization to death from any cause, while explaining that the role of other endpoints depends on disease and context. Tumor-based endpoints can support different approval pathways in different settings. The guidance does not reduce every confirmation question to a universal requirement for one particular primary endpoint. FDA's final oncology endpoint guidance
Use a simplified fictional trial with 100 participants in each of two groups. At a specified assessment, 30 in the treatment group and 20 in the comparison group meet the example's marker-response definition. The observed difference is ten percentage points. At a later fixed observation date, 55 participants in each group are alive. Those invented counts answer two different descriptive questions. The marker difference cannot be relabeled as a ten-percentage-point survival benefit.
Nor does the equal survival count establish that the treatments are equivalent. The example supplies no complete time-to-event data, uncertainty analysis, prespecified testing plan, or assessment of other outcomes. Its purpose is to show that a number has a particular meaning, not to conduct a clinical analysis from a few integers. A writer should not transform a demonstration of record interpretation into an implied treatment verdict.
Now suppose the requirement asks for durability of a response rather than only its occurrence at one assessment. A paper that reports the 30 versus 20 count may be relevant while still leaving the duration question unanswered. The analyst should identify the missing measurement rather than describe the whole paper as useless. A record can contribute part of the evidence without completing the exact regulatory task being investigated.
Cedar's evidence map therefore carries the endpoint's definition, assessment time, population, comparator, and the conclusion the document claims to support. It also distinguishes the author's numerical result from the regulator's later interpretation. The analyst can accurately report what was measured while leaving the question of regulatory sufficiency to the relevant decision record. This is not deference to a label. It is a way to make the relationship between measurement and conclusion available for criticism instead of hiding it inside the word positive.
5. Keep original milestones, revised plans and actual events separate
A fictional Cedar requirement originally projects study completion on September 30, 2024. A later plan proposes December 31, 2024. The actual completion date in the packet is December 15, 2024, and the final report is submitted March 20, 2025, ahead of a March 31 reporting milestone. The study completed 76 days after the original completion projection and 16 days before the revised projection. Its report arrived 11 days before the stated report milestone. One sentence saying on time would hide those different comparisons.
The analyst should not decide which date is legally controlling merely by selecting the latest date in a spreadsheet. The revised plan might be a sponsor proposal, an agreed amendment, or a descriptive estimate with no effect on the original obligation. Cedar's packet must identify what document changed the schedule and what authority that document claims. Without that link, the revised date is evidence of a changed plan, not automatic evidence that every earlier deadline disappeared.
FDA's postmarketing FAQ distinguishes submitted reports from fulfilled requirements and release from a requirement. It also describes delay against the original schedule, quarterly public updates, and a limited public display period for fulfilled or released records. These are requirement-level reporting definitions, not interchangeable labels for an indication's regulatory status. The FAQ's broader discussion of annual reports must not be mistaken for the separate accelerated-approval reporting provision discussed below. FDA's postmarketing requirements and commitments FAQ
For Cedar, the important finding is the sequence: projected study milestone, proposed revision, actual completion, report submission, and later agency response. A missing agency response leaves a specific question open even when the study and report dates are known. A response about one requirement does not necessarily answer another requirement for the same trial. The ledger should retain that granularity rather than inheriting a product-wide status from the first completed row.
Different milestones can also recover differently. Suppose enrollment initially falls behind a plan but later recruitment catches up before the scheduled analysis. The original enrollment delay remains part of the history; it does not prove that the final report must also be late. Conversely, finishing enrollment does not establish that analysis and reporting are finished. The article should report the event that actually occurred rather than converting progress in one stage into completion of the entire program.
None of this arithmetic alone establishes a legal violation. It identifies differences between dated events and specified schedules. A compliance conclusion would require the governing terms, applicable law, agency correspondence, and relevant circumstances. An analyst can responsibly publish a documented delay or an unresolved schedule change without pretending that a calendar subtraction supplies all of that additional evidence.
6. A trial can be underway without having answered the question
FDA's January 2025 guidance page on whether a confirmatory trial is underway remained labeled draft and not for implementation when checked for this article. The page explains that Congress gave FDA authority to require a study to be underway before approval or within a specified period afterward, as appropriate. The statutory authority and the draft's proposed interpretation are different things. A writer should not present every draft criterion as an already binding universal requirement. FDA's draft guidance on confirmatory trials being underway
Cedar's fictional trial T9 has opened five sites and enrolled 30 of a planned 300 participants. Those are three distinct numbers: sites activated, participants enrolled, and planned enrollment. Thirty divided by 300 gives ten percent of the target. It says nothing by itself about whether the trial will finish on schedule. The remaining recruitment rate, follow-up period, and operational circumstances are not supplied by that fraction.
A naive forecast might observe that 30 participants enrolled in the first three months and project another 27 months to enroll the remaining 270 at the same average rate. That is a transparent arithmetic projection of ten participants per month, not a reliable prediction. Site openings may be staggered, recruitment may accelerate or slow, and eligibility or competing studies may matter. The projection is useful only if its assumptions are visible and its uncertainty is not concealed by a precise completion date.
A different fictional plan says that the remaining sites will open next month and recruitment will rise to 30 participants per month. Under that stipulated rate, enrolling the remaining 270 would take nine months. The two projections differ by 18 months even though both start with the same observed 30 participants. The disagreement is about assumptions concerning future recruitment, not a dispute about the initial count. The record should identify what evidence, if any, supports the changed rate.
Enrollment is also not the final evidentiary milestone. If the trial needs a specified follow-up interval after the last participant enters, completing recruitment does not immediately produce the final outcome data. If the required report depends on events occurring rather than a single fixed visit, the relevant completion question changes again. Cedar does not model those clinical details; it marks them as information needed before using recruitment numbers to forecast a regulatory outcome.
A useful public update can therefore say that the study has begun and describe the verified progress without declaring that confirmation is imminent. It can compare the current record with the original plan and identify the assumptions behind a forecast. The most important distinction is between evidence that work is happening and evidence that the outstanding clinical question has been answered. The former matters, but it cannot be substituted for the latter simply because a program is under pressure to show progress.
7. Separate the reporting clock from the public observation clock
The accelerated-approval provision in 21 U.S.C. 356b requires progress reports, including enrollment progress and milestones, no later than 180 days after approval and at least every 180 days afterward until the required study is completed or terminated. It also requires prompt, searchable publication of that information. These provisions are distinct from the broader annual-report provision. The official indexed text was checked; an attempted direct House page timed out. That access limitation does not change the statutory interval. Official United States Code, section 356b
The reporting interval does not itself supply the actual timestamp for a second clock: when a member of the public can observe a particular document or status. In Cedar, a sponsor progress report bears June 24, 2025, and a researcher first observes the associated public record on July 20. The elapsed time is 26 days. That is a difference between the two stipulated dates, not proof that the agency withheld the report for 26 days or that it first received the report on June 24.
The date printed on a report, its transmission date, its receipt date, its processing date, its publication date, and a researcher's retrieval date are potentially different fields. The fictional packet provides only some of them. A responsible timeline should not silently fill the absent fields with the nearest available date. The appropriate entry for an unknown receipt date is unknown, accompanied by the specific document that would resolve it if available.
Even ordinary calendar wording deserves care. Adding 180 days to January 1, 2025, reaches June 30; adding six calendar months reaches July 1. That one-day difference does not establish an actual compliance deadline here. It shows why an analyst should preserve the interval used in the governing text instead of converting it casually into a different calendar expression. Any real deadline calculation must consider the applicable terms and circumstances, not merely reuse this example.
Suppose a search of Cedar's public folder finds no June report. Three explanations remain compatible with that observation: the report was never submitted, it was submitted but is absent from the searched folder, or it is present under an identifier the search missed. The search result alone cannot select among them. A receipt or an explicit statement about submission would change the evidence; repeating the same unsuccessful search would not automatically do so.
This is also a useful stopping rule for a public account. It can report that a specified public collection lacked a matching record when checked, describe the identifiers and dates searched, and identify which submission fact remains unresolved. It should not promote that bounded finding into a legal or clinical conclusion. The uncertainty is not vague if the missing event, relevant clock, and next evidentiary question are stated precisely.
8. Reconcile a three-document packet without granting it extra authority
Consider three additional fictional documents concerning T9. A sponsor cover letter dated May 15 calls its attachment the final report for R9. An agency receipt dated May 20 acknowledges delivery of a file bearing that identifier. A June 1 letter discusses R9 and says that a later follow-up analysis is still requested under R14. The example deliberately gives each document a narrow role. None is an actual FDA letter, and the exercise does not reproduce a real drug's correspondence.
An analyst who searches only for the phrase final report might close every requirement linked to T9. In Cedar that would affect R9, R11 and R14, because the trial identifier is shared across different indication and reporting relationships. The three-document packet does not support that operation. The cover letter explicitly names R9; the receipt identifies a delivered file; and the June letter preserves a separate question about R14. No document in this miniature packet addresses R11.
The correct first output is therefore not a single green checkmark for T9. It is a record of which claims are supported for which identifiers. Delivery of a named file is supported for R9. The existence of an outstanding follow-up request is supported for R14. The packet is silent about R11. Whether R9's evidence was sufficient for a particular regulatory outcome requires a statement that actually makes that determination, not the analyst's inference from the cover letter's title.
Now add a fourth fictional document: an amended cover sheet saying that the May attachment accidentally listed R9 when it concerned R11. This is a genuine change to the example's evidentiary situation, not a reason to erase the earlier documents. The analyst should preserve the original identifier, the correction, and the date the correction became known. A reader reconstructing what could reasonably have been concluded on May 21 should not be shown the corrected June understanding as though it had always been available.
The correction also changes the next question. Instead of asking why R11 lacks a report, the analyst must ask whether the amended attribution was accepted and what happened to R9's expected submission. That is a narrower, more productive inquiry than accusing every document of being unreliable. A record can contain a material clerical error without making every other field false; the scope of the correction must be established rather than guessed.
A publication using this packet should describe the mismatch and its consequences for the account. It can say that an earlier identifier-based summary assigned a document to the wrong requirement and that the corrected mapping changes which question remains open. It should not infer a concealed trial failure, a successful confirmation, or a motive for the mistake. Those are separate claims requiring evidence the fictional packet does not supply.
9. Read the label's version and source, not just its download date
FDA explains that current or in-use labeling in DailyMed and FDALabel can differ from the last FDA-approved labeling because some changes have not yet received approval. Drugs@FDA includes approved labeling for products regulated by the Center for Drug Evaluation and Research, but it does not cover products regulated by the Center for Biologics Evaluation and Research. The appropriate database therefore depends on both the question and the product's regulatory home. FDA's explanation of its labeling resources
FDA's 2019 final accelerated-approval labeling guidance addresses the indication's evidentiary limitations and subsequent revisions after benefit verification. It also discusses withdrawal of one indication while other indications remain approved. Its recommendations make the label a useful part of the outcome trail, not a replacement for reading the relevant decision. The document is final guidance, whereas the underway document discussed earlier is a draft; neither status should be inferred merely from the year in a search result. FDA's final accelerated-approval labeling guidance
For an original document-comparison exercise, imagine two Cedar label files. File A was approved on February 1 and downloaded on June 5. File B was downloaded on June 6 but carries no approval date in the saved excerpt. The second download is newer as an observation; it does not establish that B is the later approved version. A researcher who sorts only by local file date would answer a provenance question using a retrieval timestamp.
Suppose the files differ in three places: the distributor's address, a warning paragraph, and the wording of I2. A general statement that the label changed conceals the distinction among those edits. The researcher should locate the exact passage relevant to the investigation, preserve both versions, and identify the record connecting that passage to a regulatory event. The first two changes do not themselves establish what happened to I2, even if they make the files visibly different.
An excerpt can also create a false disappearance. If File B contains only pages one through four and the relevant indication appears on page five of File A, a text search returning no match in B proves very little. The proper comparison requires knowing whether the files cover the same sections. A cropped image, a partial download, and a complete replacement document are different objects, even when they share a product name.
In Cedar, the narrow publishable observation would be that the saved versions differ in specified text and that the packet either does or does not contain an approval record for that change. That preserves a route for later correction without pretending that file comparison alone is a clinical review. It also avoids the opposite mistake: dismissing a material change as merely a database quirk before checking what actually changed and why.
10. A snapshot's resolved cases are not its eventual success rate
Return to Cedar's 12 indications: five verified, two withdrawn, and five ongoing. Five divided by 12 is approximately 41.7 percent. If the analyst excludes ongoing indications and calculates five divided by seven, the result is approximately 71.4 percent. Both calculations are arithmetically correct. They answer different questions. The first describes the verified share of the complete fictional snapshot; the second describes the verified share among the indications already assigned one of the two resolved outcomes.
Calling the second fraction the eventual verification rate would add an assumption the data do not establish. The five ongoing indications might resolve differently, and their observation periods may differ from those of the seven already resolved. A completed-case fraction does not automatically predict the unresolved cases. The honest description keeps the denominator explicit and leaves the future outcomes unknown rather than treating exclusion as a solution to uncertainty.
A simple age split demonstrates the problem. Assign six of Cedar's indications to an older approval cohort and six to a younger cohort. The older group has four verified, one withdrawn and one ongoing indication. The younger group has one verified, one withdrawn and four ongoing indications. The older verified share is about 66.7 percent and the younger share about 16.7 percent. Those invented differences alone do not establish that the program became less effective; the younger group has had less elapsed time to reach a final outcome in this scenario.
Nor should the analyst automatically explain the entire difference by elapsed time. Different populations, requirements or other features could matter in a real comparison. Cedar supplies no data to distinguish those explanations. Its purpose is to demonstrate why a snapshot does not by itself identify the cause of a difference. A serious time-to-outcome analysis would need consistent entry dates, event definitions, follow-up information and an appropriate method, none of which is supplied by a pair of percentages alone.
The denominator can move even when no existing outcome changes. If four newly approved, ongoing indications are added to the original snapshot, the verified share becomes five out of 16, or 31.25 percent. The apparent fall from about 41.7 percent does not mean that a previously verified indication lost that status. The numerator stayed five; the population being summarized expanded. A chart that hides entry into the cohort can make ordinary additions look like reversals.
A useful public display would therefore show the count entering the cohort, the count in each outcome state, and the observation cutoff used for the comparison. For Cedar, that means explicitly reporting the five unresolved indications alongside the two resolved categories. It means not converting an administrative outcome fraction into the chance that a particular patient benefits. Those are different levels of analysis, requiring different evidence and different denominators.
11. What the local test actually checked
The accompanying local exercise used an authored JSON fixture containing the 12 indication records and 15 requirement records. A Python script counted identifiers, formed the overlapping product sets, and checked whether each requirement referenced an existing indication. The actual saved baseline result was eight products, 12 indications, 15 requirements and ten trial identifiers. It also reproduced the five verified, two withdrawn and five ongoing indication counts. These outputs concern the fixture alone, not an extraction from a government database.
The duplicate control appended a second copy of requirement R1. The resulting input contained 16 rows but still only 15 unique requirement identifiers. The checker rejected the input for a duplicate identifier instead of accepting the extra row as a new obligation. That result shows that this particular duplicate was detected. It does not prove that every possible duplicate, spelling variant or mismatched real-world identifier would be found by the simple script.
The second control replaced one requirement's indication reference with I999, an identifier absent from the fixture. The checker rejected that input for an unknown reference. Importantly, it did not invent an indication record to make the join succeed. Such an automatic repair would make the count look tidy while concealing the very discrepancy the analyst needed to investigate. In a real collection, an unmatched reference would require review of the underlying documents rather than an arbitrary new row.
The calculations also confirmed the overlapping product sets. The three outcome categories contained 11 product memberships but eight unique products. That is a useful positive check: not every count that fails to add neatly is an error. Here the overlap is a property of the authored relationships. An analyst who forces those categories to be exclusive would lose information about products with indications in different states.
There are important things this script did not test. It did not read an approval letter, determine whether a trial's population matches an indication, or decide whether a study verifies clinical benefit. It did not authenticate a signature, contact a sponsor, access patient data, or evaluate a treatment. The narrative's three-document correction exercise is a separate reasoning example, not a claim that software interpreted agency correspondence. These limits prevent a successful bookkeeping check from acquiring an undeserved scientific authority.
The script, input and output were saved separately so that the narrow result can be inspected. The fixture's digest identifies the exact bytes used; it does not establish that the invented records describe reality. Reproducibility is valuable because it lets another reader check the arithmetic and the stated failure cases. It cannot supply evidence that the input never contained, and it cannot transform an explicitly fictional exercise into a regulatory finding.
12. Write the outcome claim so a reader can see what would change it
A final Cedar account could begin with a bounded result: at the stated cutoff, the authored ledger contains 12 indication records, with five classified as verified, two as withdrawn and five as ongoing. It could then explain that those indications belong to eight products and connect through 15 requirements to ten trial identifiers. This is less dramatic than announcing how many drugs succeeded or failed, but it gives the reader enough structure to understand what was actually counted.
The next paragraph should identify the evidentiary links rather than simply repeat the totals. For I2, the fictional packet contains a study with a similar product and population name but an unresolved treatment-setting match. For T9, the correspondence exercise shows an identifier correction that changes which requirement a report concerns. Those are distinct limitations. Combining them under one general label such as insufficient transparency would make the account less useful because the missing evidence is different in each case.
Suppose a later, adequately identified decision document establishes that I2 belongs in the verified category. In the example's next snapshot, the verified count would become six and the ongoing count four, with the withdrawn count remaining two. Six of 12 is 50 percent; four of 12 is about 33.3 percent. The change would concern one indication's classification. It would not retrospectively prove that every earlier uncertainty was unreasonable or that every other indication for the product had the same outcome.
A correction should distinguish a new event from a previously missed old event. If the decision occurred after the original cutoff, the update records a genuine later development. If the decision existed before the cutoff and the analyst overlooked it, the earlier account needs a correction. The numerical change might be identical, but the explanation to readers should not be. A transparent record preserves both the event date and the date the publication learned about it.
The most useful scrutiny therefore names a resolvable question: which indication, which requirement, which study population, which endpoint, which document and which date support the claimed outcome? That question is demanding without presuming misconduct or treating approval as immunity from criticism. It makes the next piece of evidence consequential. A missing link can be investigated; a documented link can be checked; a mistaken link can be corrected without rewriting the entire story around a preferred conclusion.
Accelerated-approval follow-up deserves that precision because the public question is substantive, not merely administrative. People want to know what evidence followed an earlier decision and what the responsible agency concluded from it. A useful account must preserve the path between those events rather than replacing it with a publication count, a reassuring status word, or an alarming deadline subtraction. The goal is an outcome claim whose scope, basis and remaining uncertainty are visible enough for another person to test.