Null semantics

Null semantics is the declared interpretation of a value that is absent or marked with a special state. A blank may mean unknown, not yet researched, not applicable, withheld, or known to have no value; these are different claims. SQL NULL represents missing information in a particular data model, but a bare NULL does not identify which real-world reason applies. Codd's work on missing information made this ambiguity a central database concern.

A practical vocabulary separates states and records the evidence or policy behind each one. 'Not found' should state the search scope and date; 'not applicable' needs a reason the property does not arise; 'blocked' should identify an access limitation where disclosure permits. Missing values must survive transformations and exports without collapsing into a false negative or a default value. Query behavior, especially comparisons and aggregations, should be tested for each state.

For a defensible record, explicit null semantics prevents a gap from being mistaken for a conclusion. It also lets reviewers see whether a value could change after new evidence or access arrives. The open-world assumption warns against inferring falsity from absence in an incomplete collection, while a closed-world assumption may permit that inference inside a declared complete domain. Abstention is one response when the current null state cannot support a substantive answer.

A well-designed record separates at least five situations: not applicable, because the attribute does not exist for this entity; not yet known, because it has not been collected; unknown after search, because it was looked for and not found; withheld, because it exists but may not be shown to this reader; and not asserted by any source, because the sources are silent. Each has different consequences for a decision. Not applicable can be treated as settled. Not yet known calls for collection. Unknown after search is itself a finding that deserves a date and a record of where was searched. Withheld requires that the system say so and not suggest that nothing exists.

Represent these states explicitly, with an enumerated status beside the value or a typed marker, and not with sentinel values such as zero, an empty string or a date in 1900, which get averaged, sorted and charted as if they were data. The article on open-world systems develops the logic, and the one on knowledge graph completion shows what happens to evaluation when unknown triples are treated as false.

For each nullable field ask what a blank means and whether the schema can say. Search for sentinel values such as 0, 9999 or empty strings that stand in for unknown. Test queries that use NOT IN and joins on nullable columns. Check that data exports keep the distinction between absent and empty. Check that withheld values are labelled withheld.