A teacher opens a progress sheet. Beside one learner’s name, the result is blank. In the next column, another learner has a zero. The two cells occupy the same amount of space. They do not say the same thing.

The zero might be a recorded score. The blank might mean the learner was absent, the work has not been marked, the result was not transferred, the task did not apply, or the information is unavailable to this reader. It might also be a simple omission. None of those possibilities can be settled by looking harder at the empty rectangle.

Now imagine the sheet is passed to another teacher, imported into a database, summarised in a report and read by an AI assistant. At each step, someone may be tempted to make the record look complete. A blank becomes zero. Zero becomes poor performance. Poor performance becomes a judgement about the learner. A tiny substitution has changed the story.

The problem is not that information is missing. The problem is that the reason for its absence has disappeared too.

The answer in one paragraph

A blank is a feature of a representation, not a complete statement about the world. Before interpreting it, establish what value was expected, whether the field applies, whether collection occurred, whether the result is available, and what the source actually records about the gap. Zero, false, not found, not applicable, withheld and unknown permit different conclusions. An absence becomes useful evidence only when the observation or search process was sufficiently capable of finding what the claim says is absent.

This is the third article in a foundation sequence. What Kind of Thing Is This? separates things from claims and records. When Are Two Things the Same Thing? explains identity, copies and versions. Here we ask what to do when a necessary part of the record is missing. Read the opening distinctions for everyday use, the worked cases for practice, and the later sections for research and information-system design.

The examples below are invented teaching cases, not reports about particular learners, institutions or deployed AI systems. The proposed recording habits are an educational synthesis. Where an established technical standard is discussed, its primary documentation is linked directly.

1. A blank belongs to a record

A chair can be empty. A timetable can have an empty slot. A database can have a missing value. These are different situations, even though ordinary language calls all three empty.

An empty chair is an observed state of a physical object at a time. An empty timetable slot might express availability, an unfinished plan or an activity that has not yet been entered. A missing database value tells us something about what the database supplies. Moving from the third situation to a conclusion about the first requires evidence connecting the record to the world.

Suppose a room-booking sheet contains no reservation for Thursday afternoon. That could justify saying, “No booking is listed in this sheet.” It does not automatically justify saying, “Nobody will use the room.” A booking may be pending, held in another authorised system, or unnecessary for a particular activity. Alternatively, the sheet may genuinely be the complete booking authority. The distinction depends on the system’s rules and coverage, not the appearance of the cell.

Start with the narrowest supported description. “The value is not present in this record” is often a good beginning. Then investigate why. This prevents an observation about a document from being promoted into an unsupported claim about an event, person or organisation.

2. Zero is information, not the absence of information

A count of zero says that the count, under a stated procedure and scope, is zero. A blank does not supply that count. The difference becomes obvious when we replace abstract cells with a concrete question: how many returned books were in a particular tray when it was checked?

“The tray was checked at 4 pm and contained zero returned books” is a bounded observation. “The 4 pm entry is blank” does not tell us whether anyone checked. Both can be represented on the same sheet, but they carry different commitments.

Even zero needs context. A displayed zero might be an exact count, a rounded measurement or a value substituted by a reporting convention. A count of zero recorded submissions is not the same as a judgement that no work was attempted. A zero monetary charge might mean a service is free, a discount applies or a charge has not yet been assessed. Those interpretations require the field definition and relevant process information.

The practical rule is not “never write zero”. It is “write zero only when zero is the value established by the relevant method or explicitly defined reporting rule”. When a reporting rule assigns zero for an administrative purpose, preserve that rule. Otherwise a later reader may mistake an administrative convention for a direct measurement of the underlying situation.

3. The same blank can hide different situations

It helps to name the possibilities without pretending we already know which applies. These are working distinctions for reading and designing records, not a universal list that every institution must adopt.

Recorded situation	What it can mean	What must not be assumed
Known zero	The stated count or value is zero.	Every blank is also zero.
Unknown	The relevant value has not been established.	The answer is negative.
Not collected	The collection step did not obtain this field.	The subject lacks the property.
Pending	A defined result or process is not complete.	The final result is already known.
Not applicable	The field does not apply under its definition.	An applicable observation was missed.
Withheld or restricted	The value is not available to this reader.	Disclosure or reconstruction is authorised.
Collection or transfer failed	A step did not deliver usable information.	The underlying event failed to occur.
Not found in a search	A particular search returned no match.	No matching thing exists anywhere.
Conflicting values	Available records disagree.	The middle value is the truth.

These distinctions have precedents in real information standards. HL7’s published DataAbsentReason vocabulary distinguishes, among other cases, unknown, not asked, asked but declined, masked, not applicable and error. Its existence illustrates a design principle: a missing value and the reason it is missing need not be represented as one undifferentiated blank.

That reference is not an instruction to use a healthcare vocabulary for every educational record. Local terms need definitions appropriate to their setting. The important question is whether a future reader can distinguish situations that require different interpretations or next steps.

4. Missingness has several dimensions

A single status label can still oversimplify the problem. Imagine that a result exists, access is restricted, the source is known, and a permitted aggregate will be released later. Is the value unknown, restricted or pending? From different perspectives, all three descriptions may be relevant.

Separate three questions. Does a value exist or apply? Is a value available in this record to this reader? Why is it unavailable here? The first concerns the subject and the field. The second concerns the representation and access. The third concerns the information process. Confusing these questions encourages false explanations.

Consider a survey field that was not asked because it did not apply. The field is structurally inapplicable and no response was collected. A different respondent may have been asked an applicable question and declined to answer. Neither blank should be converted into the same personal characteristic. Their identical appearance conceals different routes to missingness.

A useful record can retain a main display label while keeping a short explanation. “Unavailable: applicable result not yet marked” is more informative than a bare dash. “Restricted: not available in this public view” says something different. When the reason itself is unknown, say so. A system should not manufacture a tidy missingness category just to avoid an untidy unknown.

5. Unknown is not false

Suppose a record does not say whether a library book has an accompanying map. There are at least three possibilities: a map is included, no map is included, or the available description does not establish the answer. Treating the third as the second makes the catalogue appear more decisive than its evidence.

The W3C’s OWL 2 Primer explains an open-world approach: a statement missing from an ontology need not be false. Some bounded information systems instead work with a closed-world interpretation for specified questions. The choice concerns what may be inferred from the representation; it does not make an incomplete record complete.

For everyday use, translate this into a question: “Is this source meant to contain every relevant instance?” A complete list of the twelve books assigned for one class can answer whether a title belongs to that assignment. It cannot establish whether the book exists, whether another class uses it or whether a teacher mentioned it informally.

There is no need to choose one attitude for the whole world. Be strict about the boundary. Within a well-defined, complete inventory, absence can answer a membership question. Outside that boundary, leave the broader question open. The disciplined reader can be decisive locally without becoming overconfident universally.

6. A failed search is a result about a search

“I could not find it” describes an outcome of a search. Its meaning depends on where you looked, what you asked for and what the search could retrieve. A complete inspection of a small labelled drawer differs from typing one phrase into a large website.

Imagine searching a school collection for a booklet whose title has changed. A query using only its old title returns no results. That observation can be correct while the conclusion “the booklet is not in the collection” is wrong. Searching the author, subject or known identifier may resolve the gap. This is why the identity distinctions in the previous article matter before absence is interpreted.

Another search may examine the correct catalogue but only public entries. A third may search a date range that excludes the item. A fourth may fail because the service did not respond. These are not four independent confirmations of non-existence. They are four different limitations of access or retrieval.

A useful negative search statement identifies the collection, the relevant date or version, the terms or identifiers used, and the important exclusions. It does not require a long technical log in the public answer. Often one sentence is enough: “No matching item was found in the public catalogue under either known title; restricted holdings were not checked.” The remaining uncertainty is now visible instead of hidden inside the word no.

7. When absence really is evidence

The warning against overinterpreting gaps should not become a rule that absence never matters. It matters when the expected thing would probably have been noticed, recorded and found under the conditions of the enquiry.

Consider an invented inventory exercise. A sealed box is opened, every item is removed, and the complete contents are counted: four blue counters and six yellow counters. The question is whether there is a red counter among those ten items. Under the stated conditions, the complete inspection supports the answer no. We are not generalising from silence in a partial description. We have bounded the collection and examined it.

Now change the example. Someone glances through a small opening and sees no red counter. The world has not changed, but the quality of the absence evidence has. A counter may be hidden behind another object. The correct claim is limited to what the inspection could reveal.

The key questions are capability and coverage. Was the method able to detect the target? Was the relevant area or period covered? Would a detection have been recorded? Would the record survive and be accessible? Was the enquiry looking for the right thing? Each weak link makes silence easier to explain without assuming absence.

This is a method for calibrating conclusions, not an excuse for endless doubt. When the boundary is clear and the observation is adequate, say what has been established. When it is not, state the narrower finding and the limitation that matters.

8. A small calculation that stops a large mistake

Use a deliberately simplified teaching model. Suppose an inspection has an 80% chance of detecting a particular defect when that defect is present. Under this assumption, the chance of missing a present defect in one inspection is 20%.

If three inspections are conditionally independent given that the defect is present, and each retains that same detection capability, the chance that all three miss it is 0.2 × 0.2 × 0.2 = 0.008, or 0.8%. The independence and capability assumptions are part of the calculation, not optional footnotes.

That result is not the probability that a defect exists after three negative inspections. It describes how likely three misses would be under the assumption that the defect exists. To calculate the reverse probability, additional information is needed, including the prior chance of a defect and the behaviour of the inspections when no defect exists.

Nor can three copies of one inspection report be substituted for three independent inspections. If all reports arise from the same observation, copying the document has not created new detection opportunities. The amount of text has increased; the observation process has not.

This arithmetic is an illustrative model, not a measured performance claim about any real inspection system. Its lesson is that the force of a negative result depends on the process generating it. Count observations, not appearances of the same conclusion. State assumptions before allowing a small probability to sound like certainty.

9. A missingness map begins before the empty cell

When a value is missing, it is tempting to start at the database. Often the more useful starting point is the route that should have produced the value. Something had to be defined, observed, recorded, transferred and presented. A gap can arise at any of those stages.

Take a fictional reading exercise. A teacher assigns a task, a learner completes it, the work is submitted, a mark is entered, an export is produced and a report displays the result. A blank on the final report does not locate the failure. Perhaps the task was never assigned to this learner. Perhaps work was completed on paper but not submitted. Perhaps the mark exists in the original record but the export excluded the column.

A sensible investigation starts by identifying the last stage supported by evidence. A submitted document supports a submission claim. A marked original supports the existence of a recorded result. Neither should be guessed from the final display alone. Trace the path only as far as the question and authority permit.

The W3C’s PROV-O specification offers a formal vocabulary for entities, activities, agents and derivation relationships. It supports recording where information came from and how one representation relates to another. It does not make a recorded value true merely because its history is documented.

For this exercise, a plain explanation may be enough: “The source contains a mark; the report export omitted it.” That is a very different finding from “No mark was ever produced.” Tracing the gap can replace an unjustified judgement about a learner with a precise description of a transfer problem.

10. The denominator decides what the number means

Missing values can change a calculation before anyone notices. Here is an invented class exercise with ten expected results. Six recorded scores are 50, 60, 70, 70, 80 and 90. Four results are unknown. The six observed scores total 420, giving an observed-score mean of 70.

Now replace every unknown with zero. The total remains 420, but dividing by ten gives 42. No new assessment evidence has arrived. The calculation has changed because the missing values were assigned numerical meaning.

Neither “the class average is 70” nor “the class average is 42” is an adequate unqualified summary of the underlying ten results. Seventy is the mean of the six observed scores. Forty-two is the result of treating four missing scores as zero. If an administrative policy explicitly requires the latter, label it as a policy-based calculation rather than an observed mean of completed assessments.

We can go further without inventing the missing values. If all ten results genuinely apply and each unknown score must lie between 0 and 100, the mean of all ten must lie between 42 and 82. The lower endpoint assigns zero to all four unknowns; the upper endpoint assigns 100 to all four. This is a logical range under stated bounds, not a confidence interval and not a prediction that the midpoint is correct.

If some missing entries are not applicable, the target calculation changes again. We must first decide whose outcomes belong in the question. A clear report therefore carries both numerator and denominator meaning: six observed results out of ten expected applicable results, with four unresolved. The missingness is part of the result, not an embarrassment to remove before presenting it.

11. Averages can hide a change in who is represented

Suppose two reporting periods each contain ten expected results. In the first period, all ten are available. In the second, only the six highest-scoring learners have submitted by the reporting deadline. A higher average among the available second-period scores does not, by itself, establish that the whole group improved.

This conclusion follows from the invented setup: the people represented changed. The comparison mixes performance with availability. A report that presents only the two means hides the most important difference between the datasets.

There are several honest options. Compare the same learners across periods while clearly limiting the conclusion to that subset. Delay the full-group claim until the required results arrive. Report available outcomes alongside coverage. Or show a bounded analysis explaining how unresolved values could affect the conclusion. Which option is appropriate depends on the decision, deadline and permitted assumptions.

The dangerous option is to delete the empty rows and forget that deletion occurred. A clean-looking table can represent a narrower group than the question originally asked about. Before comparing totals, rates or averages, ask whether the missingness changed the population being described. A numerical trend is not yet an explanation of why the trend appears.

12. SQL NULL is not a blank with a universal meaning

Technical representations need their own definitions. PostgreSQL’s official comparison documentation explains that ordinary comparisons involving NULL produce an unknown result, represented by NULL. To test whether a value is null, SQL provides predicates such as IS NULL and IS NOT NULL. A comparison such as NULL = NULL does not behave like equality between two ordinary known values.

The same system’s aggregate documentation distinguishes counting rows with count(*) from counting non-null inputs with count(expression). It also notes that, except for count, aggregates generally return NULL when no rows are selected; sum does not automatically return zero for an empty input.

These are specific documented behaviours, not rules to assume for every spreadsheet, programming language or application. They illustrate why a reader must distinguish a display, a stored value and an operation’s semantics.

For our fictional class, a count of six recorded marks and a count of ten learner rows answer different questions. Neither is wrong merely because they differ. Trouble begins when an application calls both “number of results” without explaining which count it used.

A null value also does not explain its own origin. The application may use it for an unknown value, an inapplicable field or an uncompleted transfer. Those reasons need to be retained elsewhere when they affect interpretation. Understanding the database operation is necessary, but it does not replace understanding the educational question.

13. A substitute value is a new assumption

Sometimes a calculation needs a complete set of numerical inputs. Someone may propose filling missing entries with an average, a previous value or an estimate. Such a step can be useful for a clearly defined purpose, but it creates an additional layer: what was observed and what was supplied by a method.

In the ten-result exercise, filling all four blanks with 70 produces a full-table mean of 70. It does not establish that the missing learners each scored 70. The completed table is a constructed scenario based on a substitution rule. If that distinction disappears, later users may cite invented individual results as observed facts.

Preserve the original missing status. Label the replacement as estimated, identify the rule and keep the calculation’s purpose visible. A forecast, exploratory model and official record may have different standards for what substitutions are allowed. An estimate that is acceptable for exploring a possibility may be unacceptable as a statement about a named individual.

Before selecting an estimation method, examine what would happen under alternative plausible values. If the conclusion reverses easily, the missing information is decision-relevant. If the conclusion remains unchanged throughout a defensible range, a bounded decision may be possible without pretending the values are known.

This article does not prescribe a statistical imputation method. The narrower discipline is to prevent an analytical convenience from silently becoming evidence. A completed calculation is not necessarily a completed observation.

14. A provisional value is not a missing value

A record may contain information that is available but not settled. A draft estimate, preliminary count or unverified report is not blank. It has content and a status. Replacing it with unknown can discard useful information; treating it as final can overstate it.

Consider an invented event organiser who has received twelve registration requests, confirmed eight places and not yet reviewed four requests. “Twelve registrations” may be acceptable shorthand only if the term is explicitly defined as requests received. “Twelve confirmed attendees” is not supported by these records. “The attendance is unknown” is too coarse because part of the process is known.

A better description preserves the stages: twelve requests, eight confirmations, four pending decisions. Even the eight confirmations are not observations of eventual attendance. They are records of a commitment or booking state.

Thinking in stages prevents pending information from being mistaken for failed information. It also prevents a known intermediate result from being discarded simply because the final outcome has not happened. The aim is not to maximise the word unknown. It is to describe what is known at the right level.

15. A delayed result needs more than a date stamp

A value can be missing now and available later. That does not mean a future observation proves what was true at every earlier time. Distinguish the time of the underlying event, the time it was recorded and the time a particular reader obtained access.

Suppose a fictional result was marked on Monday, entered into the source system on Tuesday and displayed in a report on Wednesday. On Tuesday morning, the report’s blank does not establish that no marking had occurred. On Wednesday, the new display does not establish that the report contained the result on Monday.

This matters when reviewing decisions. A person who acted on Tuesday may not have had Wednesday’s information. An audit should not rewrite the earlier knowledge state using evidence that arrived later. Equally, a later correction should not leave an obsolete conclusion looking current.

A useful update therefore says what changed: “The previously unavailable result has now been received; it refers to Monday’s assessment.” The earlier record can remain part of the history without being the present answer. Time makes the difference between a genuine contradiction and an ordinary sequence of information becoming available.

16. Not applicable protects the question

Some fields should not contain a value for some subjects. A question about a second book in a one-book loan does not have an unknown second title. Under the stated setup, the second-book field does not apply.

Forcing a value can corrupt both the record and the question. A form that insists on a second title may invite a user to repeat the first title or enter a placeholder. A later analyst may interpret those entries as genuine second loans. What looks like increased completeness has introduced false structure.

Not applicable also requires a rule. The label should not be used simply because the answer is difficult to obtain. In our class example, a learner exempted from an optional task may be outside that task’s result population. A learner assigned the task whose result has not arrived remains an applicable case with missing information. Those records should not be treated identically when calculating participation or results.

The useful question is: “Would a properly completed record for this subject contain a value here?” If yes, investigate the gap. If no, record the reason the field is outside scope. This small distinction helps a database describe the task that actually exists rather than the task implied by a rigid form.

17. Restricted information is not an invitation to reconstruct it

A record may intentionally omit information. That omission can be a boundary protecting a person, a confidential process or an authorised access arrangement. The existence of a gap does not establish that the reader is entitled to fill it.

Consider a fictional public report that shows aggregate learning outcomes but excludes individual names. A useful reader asks whether the aggregate supports the report’s conclusion. Trying to reconstruct the names is a different task, with different authority and privacy implications. More detail is not automatically better evidence for the question at hand.

Even the reason for withholding can be sensitive. A public system may not be able to reveal whether a particular confidential record exists. In such a case, the appropriate response can be a general statement about the access boundary rather than confirmation of a hidden item.

For an educational assistant, the safe design recommendation is straightforward: do not infer private personal attributes from gaps, do not pressure a person to answer a declined question, and do not turn restricted information into a puzzle. Seek the least information needed for the legitimate task through an authorised route. When no permitted route exists, retain the limitation in the answer.

These are responsible design principles, not a substitute for jurisdiction-specific legal advice. Their epistemic point is equally important: a guess about withheld information remains a guess, even when it sounds plausible.

18. Disagreement is not an empty space between two numbers

Suppose two records give different dates for the same event. Choosing a date halfway between them does not resolve the disagreement. It creates a third date that neither record supports.

First check identity and meaning. One date may describe an announcement and the other the event itself. One may refer to a revised schedule. A time-zone difference may place one instant on different calendar dates. The records may refer to different occurrences with similar names. Only after these possibilities are checked does a genuine conflict remain.

If the conflict remains, retain both claims with their sources and describe what is unresolved. A system can know that a disagreement exists without knowing which account is correct. That is more informative than replacing both records with a blank, and more honest than selecting whichever looks cleaner.

Where a decision cannot wait, distinguish the operational choice from the factual resolution. An organiser might use one confirmed contact channel while a records team reconciles inconsistent contact details. The fact that an action was selected does not prove that the other record was false. Good information work keeps decisions and conclusions related but separate.

19. Silence in an archive needs a history of recording

Imagine an invented collection of school newsletters. Ten issues survive from one year, and none mentions a particular club. What does that establish? It establishes an absence of mention in those ten issues, assuming they were examined adequately. It does not by itself establish that the club did not exist.

To strengthen the inference, ask what the newsletters were expected to contain. Did every issue list every active club, or only selected activities? Were issues missing? Did the club use another name? Was a separate annual register maintained? A source that never intended to record all clubs cannot be treated as a complete census of them.

Now add a complete, dated membership register that explicitly lists the active clubs for that year. If the target club is absent under all known names, the evidence for absence from that register’s defined population is stronger. It still does not settle whether an informal group met outside the register’s scope.

The example shows why negative historical claims should name their evidential boundary. “No mention was found in the surviving newsletters” is a useful result. It leaves a later researcher something to inspect and extend. “There was no such club” suppresses both the search history and the possibility that the source was never designed to answer the wider question.

20. Missing context can matter more than a missing number

A record can be numerically full and still be difficult to interpret. Consider a table containing 12, 18 and 24. Without labels, units, dates or a definition of the counted objects, the values do not answer much.

The same problem appears in a percentage. “Eighty per cent completed” needs a denominator, a definition of completion and a time boundary. Was the figure calculated from all assigned tasks, only submitted tasks or only tasks that could be checked? Are partially completed tasks included? Missing context can change the meaning even when every numeric cell is filled.

In the invented class exercise, the observed-score mean of 70 is not interpretable as a full-group mean unless the reader knows that four applicable results are missing. Here the missing information is not inside the value 70. It is in the description surrounding it.

For this reason, completeness should be evaluated against the question, not only against the number of non-empty cells. A short record with a defined subject, time, method and uncertainty can be more useful than a long table of unexplained numbers. Before asking for more data, identify which missing piece would actually change the interpretation.

21. How an AI answer can preserve a gap without becoming unhelpful

A useful assistant should not choose between invented certainty and a bare “I do not know”. There is a middle path: state what is established, identify the specific unresolved piece, explain how it limits the answer and name a proportionate way to resolve it.

For the fictional results sheet, a weak answer is “The average is 70.” Another weak answer is “There is insufficient information.” A stronger answer is: “The six available scores average 70. Four of the ten expected applicable results are missing, so this does not establish the full-group average. Check whether those results are pending, uncollected or excluded by a stated rule.”

For the catalogue search, a stronger answer is: “No match appeared in the public catalogue under the two known titles. That does not rule out a differently catalogued or restricted item. An authorised identifier-based check would narrow the uncertainty.” The answer tells the reader what the search did, rather than turning a retrieval gap into a universal claim.

These are recommended answer patterns, not claims about an existing implementation or a measured improvement in model reliability. Their value can be examined directly: the sentences keep the observation, inference and next action distinguishable. A polished paragraph should never quietly remove a limitation that its conclusion depends on.

22. Answerable questions inside an unanswerable larger question

Missing information often blocks one conclusion without blocking every useful conclusion. In the ten-result exercise, the full-group mean is unresolved. The six-result mean, the number of missing applicable results and the range implied by the score bounds are still available.

This suggests a practical reading habit: separate the original question into supported and unsupported parts. “Did everyone complete the task?” may be unresolved, while “How many completion records have arrived?” can be answered exactly from a defined dataset. Those are different questions, and the answer should not pretend otherwise.

A partial answer becomes misleading when its boundary is removed. Saying “six completed” from six received records might be too strong if the records are requests rather than confirmations. Saying “six completion confirmations are present in this report” is more precise. The extra words protect the distinction between a recorded statement and an observed outcome.

Being useful under uncertainty means finding the strongest supported sub-answer, not filling the unsupported remainder with confidence. This is especially helpful in teaching: a learner can understand exactly what additional information is needed, rather than receiving either an unjustified answer or an unexplained refusal to proceed.

23. Choosing the next question

Not every missing field deserves equal attention. Start with the decision the reader needs to make. Which gap could change that decision? Which can be resolved through a permitted, reliable route? Which is irrelevant detail?

Suppose a teacher needs to decide whether to provide a learner with the task instructions. A blank result does not first require a complex explanation of performance. A simpler discriminating question is whether the task was assigned and received. If it was never received, repeating a request for the final answer addresses the wrong stage.

Suppose instead the decision is whether to publish a group-level average. The important questions concern applicability, available results and the missing-value rule. The learner’s unrelated personal details would not resolve the calculation. Asking for them would increase intrusion without improving the answer.

A good next question divides plausible explanations. “Is the original marked work available?” separates a marking gap from a reporting gap. “Does this field apply to this task?” separates missing observation from inapplicability. “Was the search completed successfully?” separates no match from no usable search result. Ask at the point where the answer changes what should happen next.

24. Acting without pretending the uncertainty has vanished

A decision sometimes has to be made before all information arrives. That does not justify describing the missing information as known. Instead, choose an action whose justification is honest about the uncertainty.

In a low-stakes fictional classroom case, a teacher might give every learner access to a general worked example while waiting for four missing results. The action does not require identifying which of those learners needs a particular intervention. It is a provisional, reversible step with a clear purpose. A more consequential judgement about an individual would require stronger evidence.

For a report, the action might be to publish the available-score mean with coverage and limits, rather than hold every observation until the table is complete. For a claim that depends critically on missing records, the action might be to withhold that claim while releasing unrelated, adequately supported findings.

The distinction is between choosing under uncertainty and erasing uncertainty. A provisional choice should carry a reason, a scope and a condition for review. When further evidence arrives, the choice can be revisited without pretending that the earlier knowledge state was different from what it was.

25. When the outcome of an action is unknown

There is a special kind of missingness in practical work: an action was attempted, but its outcome has not been confirmed. This is neither a confirmed success nor a confirmed failure.

Imagine a fictional submission form that stops responding after a document is sent. The user sees no confirmation. The document might have arrived before the connection failed, or it might not have arrived. Repeating the submission immediately could create a duplicate; declaring failure could misdescribe the actual outcome.

The appropriate next step is to check an authorised record of the specific submission, using whatever reference the legitimate process provides. A confirmation of receipt answers whether the submission arrived. It does not necessarily answer whether the document was accepted, assessed or approved. Each later stage needs its own evidence.

This is the same grammar applied to action: attempted, received, completed and verified are different states. A missing acknowledgement is not a negative acknowledgement. The lesson is general, but the exact retry procedure should come from the system being used. The article is not granting permission to access records or repeat consequential actions outside that procedure.

26. A readable record of what is missing

A useful missingness note can be short. It should identify the question, the missing field, the observation supporting the missing status, the important scope and the permitted next step. The note need not include every detail of a technical system.

For the invented learning example: “Assessment result for task R is unavailable in the Friday report. The task applies to this learner. The report does not establish whether marking is pending or a transfer failed. Check the authorised marking record before treating the blank as a score.” This is enough to prevent several wrong inferences without exposing personal information.

For an inventory example: “No red counter was found among the ten items removed from the sealed box during the complete inspection.” This note carries a stronger negative conclusion because the scope and observation support it. It does not need to claim anything about other boxes or later contents.

The format matters less than the discipline. Separate the value from its status, preserve the reason when known, name the boundary and avoid adding a reason that was never established. A small amount of honest context can be more useful than a large collection of unexplained status codes.

27. A handoff another reader can actually use

“Please investigate” is not a complete handoff. It gives the next reader a problem without telling them where the previous enquiry stopped. A better handoff preserves the smallest useful account of the work already done.

Consider the fictional booklet search. A useful handoff says the public catalogue was searched under both known titles, no match was found, a suspected catalogue identifier remains unchecked, and restricted holdings were outside the reader’s access. The next person can continue with the identifier rather than repeat the same title searches and mistakenly count repetition as corroboration.

For the results sheet, the handoff might say that the original file was located, the learner-task relationship was checked, and the presence of a mark in the original remains unresolved. The next enquiry has a precise target. A vague label such as “data problem” would not provide that direction.

A good handoff also preserves what should not be done. Do not replace the blank with zero. Do not infer a personal reason. Do not describe a public-search result as a search of restricted holdings. These are boundaries of the evidence, not obstacles to progress. They allow another person or information system to continue without changing the meaning of what has already been established.

28. Teaching the distinction at different levels

For a younger learner, begin with objects. Show an empty transparent box, a closed opaque box and a picture of an empty box. Ask which one has been directly checked, which one remains unknown and which one is a representation. The objective is not to introduce a technical vocabulary immediately. It is to make the difference between seeing nothing and not being able to see concrete.

For a secondary learner, use the ten-result example. Ask for the mean of the available scores, then ask what is required to describe the whole group. Let the learner identify the effect of substituting zeros and derive the 42-to-82 range under the stated bounds. The arithmetic becomes a way of examining an assumption rather than merely practising a formula.

For a more advanced learner, compare a closed inventory with an incomplete catalogue, or compare missing outcome data with missing metadata. Ask for a claim that is supported, a tempting claim that is not supported, and a specific additional observation that would distinguish the alternatives.

These are suggested teaching activities, not reports of measured classroom effects. Their common objective is visible: students should learn to ask what a representation permits them to infer. A learner who can explain why an answer is not yet established has demonstrated something more useful than simply leaving the answer space blank.

29. Practice cases: decide what the gap permits

Case A: the unmarked square. A spreadsheet contains a blank beside “task completed”. The source gives no reason. Which statement is justified: the learner did not complete the task, completion has not been established in this record, or the learner declined the task?

The second statement is justified. The first and third introduce explanations not present in the evidence. A useful next question is whether a submission or completion record exists in the authorised source. The blank alone does not distinguish non-completion from non-recording.

Case B: the complete drawer. All eight objects in a drawer are removed and checked against an exact identifier. None matches. The inventory boundary and inspection are accepted for the exercise. What can be concluded?

The target item was not among those eight objects at the time of inspection. The result does not establish where it is, whether it exists elsewhere or whether the drawer had different contents yesterday. A strong bounded conclusion is better than either universal absence or unnecessary refusal to conclude anything.

Case C: the repeated report. Three websites say that no matching record was found, but all cite one catalogue search. How many independent catalogue searches have been established?

One. There are three publications carrying the result, but only one established search process. The answer could change if additional independent searches were documented. Repetition alone does not supply that evidence.

Case D: the optional exercise. Six learners completed an optional task. Four were never assigned it. Is it appropriate to describe the four as having missing scores for that task?

Not without clarifying the purpose of the record. For a table of results among assigned learners, those four are outside the applicable population. For a table of all learners, their entries may need an explicit not-assigned or not-applicable label. Neither treatment justifies inventing scores.

Case E: the later correction. A report was blank on Tuesday and showed a result on Wednesday. Does the Wednesday value prove the result was visible in Tuesday’s report?

No. It establishes availability in the later version, assuming that version was checked. Determining what the earlier report contained requires the earlier record or another appropriate account of its state. A later answer does not automatically rewrite an earlier information boundary.

30. Where this approach stops

Naming a gap does not fill it. A careful missingness vocabulary cannot repair a lost observation, establish an event that was never recorded or make a restricted source accessible. Its purpose is to prevent the gap from acquiring an invented meaning while helping the next enquiry become more precise.

The labels are also revisable. A value initially described as unknown may later be shown to be not applicable. A suspected transfer error may turn out to be an omitted collection step. Changing the label in response to evidence is a correction, not a failure of the entire approach. Keep the reason for the change visible where it matters.

There is a further limit: not every uncertainty fits a missing-value field. A complete set of observations may still support several explanations. An accurately transcribed statement may still be false. A source with excellent provenance may still be biased or mistaken. Missingness is one part of evidence-aware reasoning, alongside identity, measurement, source evaluation and the strength of the inference itself.

The objective is proportionate clarity. A simple everyday question may need one sentence of qualification. A consequential analytical claim may need a much fuller account of coverage, assumptions and unresolved alternatives. More terminology is not automatically more understanding.

Conclusion: give the gap a meaning before giving it a value

A blank can be the beginning of a useful question. It becomes dangerous when it is silently converted into a convenient answer.

Keep zero separate from unknown. Keep an unavailable value separate from an inapplicable question. Keep a search failure separate from a confirmed absence. Keep a delayed result separate from a negative result. Keep an estimate separate from an observation. Keep a privacy boundary separate from an invitation to infer.

Then ask what the gap changes. Does it alter the denominator? Does it block the conclusion or only narrow its scope? Can a permitted observation resolve it? Is a reversible next step possible while the answer remains open? Those questions turn missing information into a manageable part of reasoning instead of an empty space to conceal.

Before filling the blank, establish what the blank stands for. Before concluding that something is absent, establish what would have revealed its presence.

Sources, scope and further reading

The external references support the technical distinctions indicated in the article, not every proposed classroom activity or fictional example. Consult the W3C OWL 2 Primer for the open-world distinction; W3C PROV-O for provenance relationships; the cited HL7 terminology release for explicitly differentiated absence reasons; and PostgreSQL 18 documentation for NULL comparisons and aggregate behaviour. Software-specific examples should be checked against the version actually in use.

Continue the foundation sequence through types, evidence and claims and identity, copies and versions. For the wider management of records, see Data Quality. This article’s narrower concern is what a particular gap permits a reader to conclude.

eduKate Publishing. Educational explanation with fictional worked examples; not a claim of deployed AI functionality or a measured validation result.