As-of query

An as-of query asks for a record at a specified time rather than simply returning its latest state. The question is incomplete unless it says which clock it uses. A valid-time query asks what applied on the target date according to a chosen knowledge state. A transaction-time query asks what the system held at the target date. A bitemporal query can state both dates and expose the difference.

In practice, the query contract should define time zone, precision, interval boundaries, and how open or unknown dates behave. A replay must include the relevant identity links, classifications, and source versions, not only one field's old value. A change to an interpretation rule can also alter a historical answer unless the rule version is retained. The output should show the two temporal coordinates when they matter to the reader.

Organizations use as-of queries to examine what supported a past decision or to reassess an earlier date under today's evidence. Those are distinct exercises with different answers, both potentially legitimate. Naming the time axis prevents a retrospective correction from masquerading as contemporaneous knowledge. The query is a retrieval operation; point-in-time reproducibility is the broader ability to recreate the entire decision context and explain its inputs.

An as-of answer is only as good as the history behind it. Check three things. First, coverage: the store must have been recording at the date asked about, and a query for a date before the system began should return a clear statement of that, not an empty result that looks like absence of facts, a case of the closed-world mistake. Second, resolution: if changes were batched, the answer is correct only at batch boundaries. Third, inputs: if the answer depends on retained sources, those sources must also be retrievable as of the date, which links the query to retained source copies.

As-of queries are also a testing tool. Replaying a decision with the data as of the decision date tests whether the stored inputs were sufficient to reproduce it. The article on temporal knowledge graph benchmark leakage shows the converse problem: evaluating a model with data that were not available at the time of the prediction, which as-of discipline prevents.

Run an as-of query for a date before and after a known correction and compare the answers. State which time axis the query uses. Query a date before the system started and check the response. Confirm that results carry the as-of parameters in their output, so that a printed report shows the date it describes. Confirm that the sources behind an old answer are still retrievable.