1. A case total answers a narrower question than a claim about the sky
A table of unidentified anomalous phenomena can look deceptively decisive. It has a total, several categories and perhaps a percentage left unresolved. The arithmetic may be correct while the interpretation is not. A proportion of submitted cases that remains unidentified is not automatically the probability that an unfamiliar object is extraterrestrial, the frequency of unusual events in the atmosphere or the chance that a particular witness is mistaken. Those are different questions, involving different populations and different evidence.
The fiscal 2024 AARO report provides a dated example. It records 757 reports received during its reporting period, of which 485 concerned incidents in that period and 272 concerned incidents from 2021–2022. Those numbers describe receipts and occurrence dates, not 757 newly occurring events in the report's publication year. This article uses that historical report as a bounded example, not as an October 2026 current-case census. AARO fiscal 2024 consolidated report.
The first obligation is to identify what was counted. A report may describe an observation. Several reports may concern one episode. A case may combine several observations. A public file may contain a report, an analyst's note and a later correction. Counting any one of those units can be useful, provided the account does not silently rename it as another. A file count can measure the size of a document release without establishing the number of independent events represented in it.
To make the distinctions testable, this article uses a wholly fictional program called Harbor. Its records are invented for arithmetic and evidence-reconciliation exercises. They are not reconstructed AARO cases, real witness statements, military flight logs or a simulation calibrated to actual UAP behavior. The local checker counts the supplied rows, applies an explicitly defined status update and rejects several corrupted inputs. It neither identifies objects nor assigns probabilities to extraordinary explanations.
The useful outcome is a better reading of the evidence already available. An analyst should be able to say that a database received more reports, that a particular group of cases was resolved or that a public release contains a certain number of files. Each statement can stand on its own without becoming a claim that the whole sky has changed. Where the data cannot distinguish explanations, the right result is an identified limitation, not a percentage invented to fill the gap.
2. Separate reports, episodes, witnesses and files
NASA's Aviation Safety Reporting System illustrates why database units need documentation. Its public explanation says that multiple participant reports about one incident can be combined into a single record. The same explanation cautions that voluntary reports are subject to self-reporting biases, are not verified by NASA ASRS and cannot establish the prevalence of a problem in the National Airspace System. These are stated limits of that particular database, not a claim that every UAP collection uses the same processing rules. NASA ASRS: About the data.
Harbor's supplied fixture contains twenty case identifiers, C01 through C20. Each initially has one report. Four additional reports refer to C01, C02, C03 and C04. The file therefore contains twenty-four reports linked to twenty cases. Neither total is an error. The reports table measures submissions, while the cases table measures the episodes grouped under the exercise's stipulated identifiers. The four additional reports do not create four more episodes merely because they occupy four more rows.
Nor does this fixture establish a witness count. It deliberately has no verified person-identity table. One person could have submitted more than one report, or a report could summarize accounts from several people. Calling the twenty-four submissions “twenty-four independent witnesses” would add information that the exercise does not contain. Independence is not supplied by a distinct file name, a second upload or a new account repeating a description first encountered elsewhere.
The actual checker verifies that every report points to an existing case identifier. When one report is deliberately changed to point to C99, it rejects the dataset because C99 is absent from the case table. It also rejects a duplicated case identifier. Those controls demonstrate limited relational consistency. They cannot determine whether the original decision to combine two real-world observations into one case would have been justified. That decision would require the underlying timing, location and evidentiary record.
A defensible public description could say that twenty cases are represented by twenty-four submissions under Harbor's supplied grouping. If the grouping is uncertain, preserve that uncertainty. An analyst might maintain a proposed match without collapsing the records, or show totals under two explicit grouping assumptions. The important thing is to make the transformation visible. Once the original submissions are replaced by an unexplained case count, readers cannot tell whether a later change reflects a new observation, a correction or a revised judgment about which reports belong together.
3. Keep the occurrence calendar separate from the receipt calendar
All twenty Harbor cases were received in the exercise's 2026 intake year. Six concern occurrences assigned to 2025, and fourteen concern occurrences assigned to 2026. A chart labeled “reports received in 2026” would therefore show twenty. A chart labeled “2026 occurrences among the supplied cases” would show fourteen. The six-case difference is explained by the fictional dates, without requiring any change in the number of events that actually occurred in either year.
This distinction becomes especially important when comparing successive releases. Imagine that Harbor opens a new submission channel and receives ten older accounts in one week. Its receipt workload rises immediately. That increase is not, by itself, evidence that ten new events happened that week. A separate occurrence-date view could reveal that the accounts concern several earlier years. The receipt chart still matters for administration and public interest, but it should not be presented as a direct time series of physical activity.
A case can have more than two relevant dates. Consider an invented observation on December 28, a report submitted January 3, a file accepted January 5 and an analytical decision recorded March 12. Different dashboards might legitimately assign it to different periods, depending on whether they track observation, submission, intake or completion. A reader should ask which date drives the count. If the time zone or the meaning of a timestamp is unknown, the chart should not imply more precision than the record supplies.
The difference also affects claims about sudden clusters. A dozen files released together can reflect a publication batch rather than a dozen simultaneous occurrences. The release date is an event in the history of the records. It does not replace the date of the reported observation. To investigate a physical cluster, one would need the occurrence times and their uncertainty, together with evidence that the observations refer to distinct episodes rather than repeated accounts of the same one.
Harbor's six older cases should remain visible in its 2026 receipt reconciliation. Removing them from that table would understate what the program processed. They should also remain identifiable as older occurrences so that a reader can construct the relevant historical view. The solution is not to choose one calendar and discard the other. It is to retain both, name each measure and explain why the same set of records can produce different annual totals.
4. Classification labels are decisions within a defined system
GEIPAN's classification explanation distinguishes cases with an accepted or probable explanation, cases that lack reliable information for assessment, and unexplained cases assessed using its criteria for information quality and strangeness. It also permits reclassification when new information arrives. Its categories are not interchangeable with every other organization's labels, and this article does not treat the older numerical totals on that explainer as current October 2026 statistics. GEIPAN classification methodology.
Harbor uses a different, invented four-state scheme. At the first snapshot, eight cases are marked resolved, eight data-limited, three under active analysis and one queued for initial work. These categories sum to twenty because the fixture assigns exactly one current workflow state to each case. The labels are stipulated administrative inputs. They are not independent conclusions reached by the local script, and they do not describe a real government's investigative standards.
The distinction among the twelve cases not marked resolved is useful. The queued case has not yet reached the same stage as the three under active analysis. The eight data-limited cases lack the information required by the exercise's decision process. Combining all twelve into an “unresolved” group may answer a broad workflow question, but it removes those distinctions. It would be especially misleading to rename all twelve “cases that defeated detailed investigation” when the supplied states establish no such common history.
A single case may also have several different attributes. C17 can be under active analysis, publicly released and linked to an observation at night. These fields are not competing alternatives. A pie chart should not add “under analysis,” “public” and “nighttime” as if every case belonged to exactly one of those groups. Before making percentages sum to one hundred, establish whether the categories are mutually exclusive and whether they exhaust the defined population.
The checker includes a deliberately invalid status called “confirmed_alien” in a corrupted copy of the fixture and rejects it because it is outside Harbor's defined vocabulary. That is a schema test, not a scientific rejection of a hypothesis. The program would reject any unsupported label in the same field. A successful validation tells the reader that the supplied rows follow the chosen vocabulary; it does not establish that the real-world interpretation behind each row is correct.
5. Do not divide all completions by only the new arrivals
Harbor resolves eight of its twenty newly received cases during the fictional reporting period. It also resolves five cases inherited from earlier periods. Its total completed workload is therefore thirteen cases. The resolution fraction within the new intake cohort is eight divided by twenty, or forty percent. Dividing all thirteen completions by the twenty new arrivals produces sixty-five percent, but that ratio does not mean sixty-five percent of the new cases were resolved.
The mixed ratio can still have a defined administrative use if clearly labeled. It compares total completions with new intake during the period. What it cannot do is substitute for a cohort outcome. The five older completions belong in a workload account and in the outcome history of their original cohorts. Adding them to the numerator while excluding their cases from the denominator changes the question. No amount of decimal precision repairs that mismatch.
The unfinished-work balance makes the relationship concrete. Harbor begins with thirty unfinished cases, receives twenty and completes thirteen. Assuming no transfers, removals, mergers or reopenings, it ends with thirty-seven unfinished cases. The checker reproduces that balance. The backlog grows by seven even though thirteen cases were completed. A statement that the office “did nothing because the backlog grew” would ignore completed work; a statement that it “cleared the backlog by resolving thirteen” would ignore the incoming work and remaining total.
Now introduce a different event: two duplicate cases are merged after intake. That would require another term in the reconciliation, because the count of case records changes without a new object being identified. The original thirty-seven balance would no longer be complete under that altered scenario. The report should explain the merge and identify the affected records rather than force the new ending total to match an equation whose assumptions have changed.
A public table should therefore preserve at least the opening population, additions, completed cases and other count-changing actions. It should also distinguish new-cohort outcomes from older-cohort outcomes. These are not demands for secret operational information. They are definitions needed to understand published totals. Where disclosure limits prevent a complete reconciliation, the limitation should be stated instead of presenting a partial balance as though every difference had been explained.
6. A solved subset need not represent the unsolved subset
NASA's September 2023 independent study recommended more systematic UAP data collection and described reporting stigma as an obstacle to obtaining information. These were recommendations and observations about the collection problem, not a census of every event or a validated probability model for unresolved cases. The distinction remains important when using the report as methodological context. NASA UAP Independent Study Team final report.
Harbor's fixture assigns eight cases to a relatively richer-information group and twelve to a sparse-information group. Six of the richer cases are resolved, compared with two of the sparse cases. The within-group fractions are therefore seventy-five percent and about 16.7 percent. Across all twenty cases, the fraction is forty percent. These are authored inputs and calculated results, not measurements of any agency's ability or evidence about the actual causes of UAP reports.
The exercise shows exactly what an extrapolation would assume. If an analyst takes the resolved group's composition and assigns it to every unresolved case, the analyst assumes that the relevant differences between the groups do not affect that composition. The fixture supplies no basis for that assumption. Information richness is already associated with completion in the invented data. That does not prove any specific alternative explanation; it establishes why the proposed transfer needs an argument beyond the fact that some cases have been solved.
The reverse extrapolation is equally unsupported. The existence of harder-to-resolve cases does not establish that they involve a fundamentally different kind of object. A missing timestamp or absent follow-up record could prevent identification of an otherwise ordinary event. A richer record might still leave a serious question unanswered. The classification states alone do not distinguish all those possibilities. A reader should resist both the universal dismissal and the universal extraordinary interpretation when neither follows from the supplied evidence.
A second invented selection exercise makes the reporting stage visible. Assume one hundred observable episodes of a chosen kind in each of two periods. If ten are reported in the first period and twenty in the second, received reports double while the stipulated episode count stays constant. This is not an estimate of actual reporting probabilities. It is a counterexample showing that a doubled receipt count does not logically require a doubled occurrence count. Real trend analysis would need evidence about what changed in observation, reporting and processing.
Improved reporting can be valuable even when it complicates a trend chart. More accounts may create opportunities for corroboration and follow-up. The analyst's job is not to dismiss that improvement as meaningless, but to measure it honestly. A record of increased reporting is a finding about reporting. Establishing a change in the underlying phenomenon requires additional information rather than a more confident adjective attached to the same count.
7. Exposure is a measurement, not a convenient large denominator
FAA's OPSNET documentation distinguishes airport operations, which cover arrivals and departures and exclude overflights, from tower and other facility measures. Those are specific operational counts, not a universal measure of every opportunity to observe an unusual object. Borrowing a convenient aviation total does not automatically create a valid denominator for an unrelated set of reports. FAA OPSNET overview.
In Harbor's entirely separate toy exposure table, the North station supplies twenty reports during one thousand logged observation hours. The South station supplies five reports during one hundred hours. North has four times as many reports, but its supplied rate is two reports per hundred hours, compared with five at South. The ranking reverses when the stipulated observation time is considered. Neither ranking is false; the raw count and the time-normalized rate describe different aspects of the invented records.
That rate is still not a probability per independent event. A report can occur more than once in an hour, and the fixture does not define a set of independent trials. Nor does it establish that an hour at North offers the same conditions as an hour at South. The sites might differ in visibility, equipment coverage, reporting practice or the activities present. The calculation is a descriptive rate under stated inputs, not a causal finding that one location attracts more extraordinary phenomena.
The time log itself needs checking. Suppose North's one thousand scheduled hours include two hundred hours when the instrument was offline. If the intended denominator is operational monitoring time, the relevant value becomes eight hundred, and twenty reports correspond to 2.5 per hundred hours. If observers continued making reports while the instrument was offline, a different measure may be needed. The record should state what the hours represent rather than assume that scheduled, staffed and instrument-operating time are interchangeable.
A duplicate time interval can cause a similar problem. Two overlapping files may each describe the same two-hour shift. Adding their durations as though the periods were disjoint doubles that portion of the exposure without adding observation time. The analyst should inspect interval identity before summation. This is a data-reconciliation task, not a claim that every monitoring program has made such an error.
The useful public result would retain the raw counts, the defined exposure and the limitations of comparison. It could report that South's rate is higher in this particular table while refusing to infer the reason. A careful denominator makes the arithmetic more informative, but it does not replace the evidence required for a causal or physical explanation.
8. A public release is another selected population
NARA's UAP collection FAQ describes ongoing, rolling transfers and distinguishes records available in the online catalog from material accessible through other arrangements. It also points readers to older UFO and UAP-related holdings across multiple record groups. A public collection is an evidentiary resource, not an automatic one-record-per-event registry or a declaration that every account it preserves is true. National Archives UAP collection FAQ.
Imagine a separate Harbor publication exercise with one hundred case records: seventy resolved and thirty not resolved. Its editors post ten resolved cases and twenty unresolved cases. In the complete invented case table, the unresolved share is thirty percent. Among the thirty posted cases, it is about 66.7 percent. The larger public percentage follows from the chosen publication set. It does not require the underlying case population to have changed.
There could be many reasons for a selected release. Some files might be easier to prepare, already public, especially informative or affected by restrictions. The toy exercise does not assign a motive, and a real analyst should not invent one from the observed difference. What can be established is the selection rule if documented, the set actually inspected and the limits of an inference from that set to material not inspected.
File multiplicity introduces another layer. Suppose one published case is represented by a narrative PDF, three photographs and a correction notice. That is five digital objects concerning one case under the exercise's grouping. Another case may have only a single summary. A count of downloaded files would then weight the first case five times as heavily if mistakenly treated as a case count. A useful manifest connects each object to its role and to the case grouping being used.
Consider an analyst who downloads twelve objects but has a manifest listing fifteen. The immediate conclusion is that three listed objects are missing from the analyst's retained set. It is not that three observations were concealed, three cases are unresolved or three records were destroyed. The missing objects might be alternative formats, a failed download or material unavailable in that channel. The account should identify the actual discrepancy and the evidence needed to resolve it.
A release can still be important without being a representative sample. It may document a particular decision, permit scrutiny of a disputed explanation or reveal the history of an investigation. Those are worthwhile uses. The statistical mistake is to turn a deliberately or incidentally selected public set into a census of all observations, without showing how the selection occurred or why the proposed generalization is justified.
9. Missing descriptions change the range of possible percentages
Harbor's twenty-case fixture records a usable shape description for only eight cases. Four of those eight are described as round. Within the known-description subset, the round share is fifty percent. Across all twenty cases, however, the supplied information does not establish that half were round. If none of the twelve missing descriptions would be round, the share is four out of twenty, or twenty percent. If all twelve would be round, it is sixteen out of twenty, or eighty percent.
Those endpoints are simple bounds under the invented premises, not a confidence interval and not an estimate of where the answer probably lies. They identify what remains compatible with the missing information as defined. Choosing the midpoint would add an assumption; labeling the endpoints a statistical uncertainty band would imply a procedure not performed here. The wide range is informative precisely because it shows how little the known subset alone establishes about the whole table.
The missing field also needs interpretation. A description may have been unrecorded, unavailable to the public, illegible or intentionally left blank because the observer could not determine a shape. Those reasons have different evidentiary consequences. An analyst should not convert every blank into “no visible shape,” nor treat a withheld value as a random draw from the values disclosed. A separate missingness code can preserve the distinction without exposing restricted or identifying information.
Descriptions need not be exclusive. In an additional invented ten-case table, six cases receive a “light” tag and five a “round” tag, with three receiving both. The number receiving at least one of those tags is eight, not eleven. Two receive neither. Adding the tag totals would count the overlap twice. A chart can show multiple-response categories, but its caption should explain that the percentages need not sum to one hundred.
Nor does a word establish a physical classification. A record containing “round” preserves a description in a particular context. It does not independently establish a spherical three-dimensional object, its size or its origin. Those questions require other evidence. This article does not resolve them by relabeling the text. It keeps the observed or reported description separate from the inference someone might later want to draw.
The strongest summary would therefore state both the known-description denominator and the extent of missing information. A reader can then understand why a chart among described cases differs from a claim about all cases. Excluding missing rows may be appropriate for a clearly named descriptive question; concealing their exclusion is what makes the result appear more comprehensive than it is.
10. Freeze a snapshot before comparing changing statuses
Harbor's second snapshot changes C09, C10 and C17 to resolved. The first two had been data-limited; the third had been under active analysis. The updated totals are eleven resolved, six data-limited, two under active analysis and one queued. The case population remains twenty. The fraction marked resolved rises from forty to fifty-five percent, a fifteen-percentage-point increase, without the receipt of a single new case in this particular update.
The change needs an explanation record. Perhaps the exercise supplies a new observation, corrects a join or records a decision that was pending at the previous cutoff. The status table alone does not identify which reason applies. An analyst should retain the old state, the new state, the effective date and the supporting reference. Overwriting yesterday's table without a change log makes it difficult to distinguish actual progress from a corrected data entry.
Real published documents can also contain different status snapshots. The fiscal 2024 AARO report's executive summary describes 118 resolved cases and 174 additional closures subsequently finalized. Its later analysis section presents 49 resolved and 243 recommended for closure. Both pairs sum to 292, but the partitions differ. A reader should preserve their stated wording and dates rather than silently choosing one partition as a timeless result. The document alone should not be made to explain an unstated reconciliation. AARO fiscal 2024 report, executive summary and analysis.
The same caution applies to newer material. An official indexed AARO congressional-products listing identifies a fiscal 2025 report added July 20, 2026. This article did not obtain its body through the directly accessible listing and therefore does not reproduce third-party claims about its totals. A newer title is not permission to combine unverified numbers with an older, directly read report. AARO congressional and press products.
A snapshot comparison should first confirm that the same case population is being compared. If five new cases arrive between Harbor's tables, the new denominator is no longer twenty. If a duplicate is merged, the population changes for another reason. If an old case is reopened, the resolved count may fall without any observation being erased. These changes are not inherently suspicious; they require different terms in the reconciliation and different wording in the account.
The publication should retain a cutoff and a version identifier. That permits a later correction to say exactly what changed, instead of implying that anyone who quoted the earlier accurate snapshot must have been dishonest. Dynamic records can support careful reporting, but only if their history remains visible enough to reconstruct the comparison.
11. Test the claim at the level where the evidence exists
Suppose Harbor's C17 packet contains a witness description, an initial analyst note and a later decision. Each has a different author and function. The witness describes what seemed unusual. The analyst proposes an explanation conditional on a time match. The later decision accepts that explanation after the exercise supplies a corrected timestamp. A public account should distinguish those stages rather than quote the earliest uncertain phrase as though it were the final finding or the final finding as though it had been present from the start.
The corrected timestamp itself should be traceable. Imagine the first file says 21:14, while a retained clock log supports 21:41. Those digits are similar but describe different times. If the proposed explanation depends on the change, the account should identify the basis for it. A typographical correction supported by a source record is not the same as changing the time merely to improve a match. The exercise does not say which event actually occurred in any real sky; it shows what the documentary chain would need to establish.
An explanation can also be conditional rather than categorical. If the analyst writes that a particular match is plausible provided the revised time is correct, removing that condition changes the claim. Conversely, a critic should not portray the condition as proof the explanation is false. The right next question concerns the evidence for the time. Disagreement can be narrowed to a testable proposition instead of expanded into a contest over whether every investigator or every witness is trustworthy.
The absence of a supplied original record is another limit. If a public packet contains only an analyst's summary of a video, the reader can evaluate the summary as a document but cannot honestly claim to have independently inspected the video. A quotation of an official conclusion is evidence that the conclusion was stated. It is not a substitute for reproducing the underlying analysis, and the account should say which kind of work it actually performed.
A practical evidence note could list the documents inspected, the calculation reproduced and the unresolved dependency. For C17, that might mean the timestamp correction is documented but the original image sequence is not supplied. Such a note is more useful than either an unqualified endorsement or a blanket accusation of concealment. It tells another reader where independent checking can begin and where the present account reaches the boundary of its access.
The same standard applies to the synthetic exercises here. Their inputs and checker results are retained, and their arithmetic can be recomputed. That establishes reproducibility of the invented examples. It does not turn them into observations, validate a real agency's classifications or prove a conclusion about the origin of any actual report.
12. An investigation priority is not a discovery verdict
Harbor's separate triage exercise contains one hundred invented cases. A first-stage rule flags twenty for prompt review. In the hand-assigned comparison labels, eight of the flagged cases need urgent attention and twelve do not. Among the eighty unflagged cases, two need urgent attention and seventy-eight do not. These values are stipulated to examine a decision rule. They are not a measured screening system for UAP or a recommendation for handling real aviation hazards.
The fraction of flagged cases needing urgent attention is eight divided by twenty, or forty percent. The rule captures eight of the ten urgent cases, or eighty percent, and misses two, or twenty percent. Those three percentages answer different questions. Calling the rule “eighty percent accurate” would hide which question was measured. The total of matching classifications is eight plus seventy-eight, or eighty-six out of one hundred. The arithmetic must be checked rather than inferred from an impressive-sounding percentage.
The costs of errors are not supplied by those counts. Missing an urgent case could matter differently from spending time reviewing a case that turns out not to be urgent. A decision-maker would need the actual context and consequences. The toy table cannot identify an optimal threshold, and it does not show that a flagged case is extraordinary. It shows that a workflow label can be designed to allocate attention rather than to announce a final physical conclusion.
A changed threshold would require new results, not a relabeling of the old ones. Suppose an imagined second rule flags thirty cases, capturing all ten urgent cases but also twenty non-urgent cases. It has no missed urgent cases under these stipulated labels, but only one-third of its flagged cases are urgent. Whether that trade-off is preferable depends on resources and consequences absent from the exercise. The example is not an operational instruction; it exposes the information a claim of “better screening” would need.
The comparison labels can themselves be uncertain in a real investigation. Here they are assigned so that the arithmetic is unambiguous. In practice, an unresolved case cannot automatically serve as a definitive positive label for an extraordinary cause. Training or scoring a system against that substitution would measure agreement with a workflow state rather than discovery of the underlying phenomenon. The account should identify the target being predicted before reporting a performance number.
This distinction protects serious investigation. A case can deserve attention because of a safety concern, a potentially useful observation or an unresolved evidentiary question. None of those reasons requires an advance verdict about its origin. The public can evaluate priorities more intelligently when the criteria for allocating work remain separate from the conclusions that work may eventually support.
13. Publish a count that can survive its next update
A concise Harbor summary could state that the first snapshot contains twenty cases represented by twenty-four submissions, with eight cases resolved under the supplied workflow labels. Fourteen of the cases concern occurrences assigned to the intake year; six concern older occurrences. Five additional completions came from the inherited workload and are not included in the new-cohort resolution fraction. That description conveys several useful results without pretending to estimate the frequency of unidentified objects throughout the atmosphere.
The second snapshot could then report three status changes, increasing the resolved count to eleven while preserving the same twenty-case population. It would identify the changed cases and explain that the updates do not create new occurrences. If the actual supporting decisions were unavailable to a public reader, the summary would distinguish the recorded changes from independently verified explanations. Clear provenance is compatible with reporting an official or supplied classification; it simply prevents the attribution from disappearing.
The retained local exercise checks counts, linked identifiers, status vocabulary and arithmetic. Its negative controls deliberately introduce a duplicated case, an unknown case link and an invalid status. All three are rejected. The additional tables demonstrate publication selection, missing-description bounds, exposure-normalized rates and triage trade-offs. These checks have a defined scope. They do not inspect aircraft, contact witnesses, operate sensors or evaluate classified information, and no such activity is implied by calling the calculations reproducible.
A useful correction policy follows from the same design. If R21 is found to duplicate another submission rather than describe a separate report, update the report count without automatically deleting its linked case. If two cases are later shown to describe one episode, update the case grouping and explain which previous totals change. If a missing shape description is recovered, recompute the known-subset percentage and the remaining bounds. Each correction should identify its effect instead of silently replacing the entire history with the latest number.
Readers should also be able to distinguish a failed check from an unanswered scientific question. An impossible foreign key is a detectable data error. A case without enough evidence for identification is a limitation of inference. A proposed explanation that contradicts reliable observations is a substantive problem for that explanation. Treating all three as the same kind of “mystery” prevents the appropriate next step from being identified.
The central discipline is to keep the denominator attached to the claim. Counts of reports, cases, public files, completed investigations and observation hours can all contribute to understanding. None becomes a probability of an extraordinary origin merely by being divided by another available number. A careful account can acknowledge unresolved observations, value witnesses, scrutinize official reasoning and still refuse to manufacture certainty where the supplied records do not support it.