Is a certain answer always explicitly stored?

No. It may follow from constraints or inference rules.

Is this just SQL's treatment of NULL?

No. SQL NULL semantics and certain-answer semantics are related topics but are not automatically equivalent.

Can certainty change when new information arrives?

Yes. The admissible interpretations and relevant assumptions can change, altering what follows from the database.

AI Glossary / Data engineering and architecture / Intermediate

Certain answer

Definition

A certain answer is a query result that holds in every admissible interpretation or completion of an incomplete database. It distinguishes what the available information guarantees from what might be true in only some possible situations. Certainty is always relative to the stated facts, constraints and rules defining those possible situations.

What it means

A database can leave facts unknown rather than false. For example, a person's department may be missing, or several records may be consistent with more than one real-world arrangement. A possible answer holds in at least one allowed completion; a certain answer holds in all of them. This is a logical notion of guarantee, not a confidence score from a statistical model. It also does not certify that the recorded facts themselves are correct.

How it works

A formal representation specifies the set of complete databases compatible with the incomplete information. Conceptually, evaluate the query in every such database and retain answers common to all results. Actual systems try to derive those answers without enumerating every completion, which may be enormous or infinite. The query language, constraints and open- or closed-world assumptions affect both meaning and computational difficulty. Some research settings require richer answer representations than a simple intersection of returned tuples.

Three practical examples

A data-integration system returns only customer-region assignments supported under every allowed resolution of missing fields.
A knowledge-based application uses rules to infer query results that follow regardless of how other unknown facts are completed.
An AI data-quality study distinguishes conclusions stable across plausible missing-data completions from predictions that depend on one arbitrary imputation.

Think of it this way

Think of several maps that agree on the roads already surveyed but differ in unexplored areas. A statement that remains true on every allowed map is certain relative to that collection. The analogy has limits: the allowed completions are defined mathematically, and excluding a legitimate map by mistake can create a misleading appearance of certainty.

Worked example

Suppose a table records Alice in Sales and Bob with an unknown department constrained to either Sales or Support. Query: “Which people are in Sales?” One admissible completion returns {Alice, Bob}; the other returns {Alice}. Their common answer is {Alice}, so Alice is certain and Bob is possible but not certain. Now query: “Does Bob belong to either Sales or Support?” Under the stated constraint, the answer is yes in both completions, so that proposition is certain despite the missing department. If the allowed departments also included Engineering, the second conclusion would change. The example shows why constraints are part of the meaning.

Illustrative case study

Fictional case: Elm Integration combines two departmental systems and finds missing assignments for some employees. A prototype fills every missing value with the most frequent department and presents the resulting report as definitive. The team instead separates guaranteed memberships from memberships that depend on an imputation. It explains the allowed values and highlights unresolved records for review. This fictional case illustrates a change in information handling, not a claim about employment decisions or a production database's performance.

Common misconceptions and mistakes

Do not interpret a missing value as evidence that no value exists. Do not equate certain with highly probable: an answer with 99% probability can fail in an admissible completion and therefore be uncertain in this logical sense. Conversely, a logically certain answer depends on the correctness of the underlying assumptions. Query negation, aggregates and inconsistent data require particular care. If there are no admissible worlds because the constraints contradict each other, a naive “true in every world” interpretation can become vacuous; the system must define how inconsistency is handled.

Business use

Businesses can use the concept to communicate what incomplete integrated data actually support. It is useful when a report should distinguish guaranteed findings from tentative ones. The trade-off is that conservative answers may omit useful possibilities, and exact computation may be expensive. Decide whether the application needs certain answers, possible answers, probabilities or explicitly labeled approximations. For AI workflows, do not present a single imputed dataset as if it resolved all uncertainty about the source.

Check your understanding

A record's category can be A or B, and a query asks whether it is A. Is A a certain answer because it is allowed?

Answer: No. It is possible, but the B completion prevents it from being certain. A guarantee requires the answer to hold in every admissible completion.

Related terms

Possible world
incomplete database
possible answer
open-world assumption
constraint
SQL NULL
imputation

Learn next

Construct two or three allowed completions of a small incomplete table and evaluate the same query on each. Compare the intersection and union of the results. Then change one constraint and observe how both sets change, making the role of assumptions explicit.

Sources

Primary research reference: Libkin, Certain Answers as Objects and Knowledge, author-hosted manuscript, https://homepages.inf.ed.ac.uk/libkin/papers/aij.pdf ; related primary research, https://arxiv.org/abs/1904.00234 . Author manuscript opened and research search excerpts consulted; full semantics-specific and editorial review remain pending.

Previous: CERNO test Next: Certainty effect