How should an earlier AI-assisted decision's historical accuracy be distinguished from its continuing validity today?

Direct answer

Assess every decision against at least two time questions. Historical accuracy asks whether the result was supported by the facts, sources, rules and authority at the time of decision. Current validity asks whether the decision remains active today and has not expired, been suspended, withdrawn, corrected or replaced. An earlier decision may continue to produce effects even after becoming invalid; those effects require separate detection and remedy.

In plain language

A driving licence that was valid last year was genuine at the time. If it has now expired, that does not make the old document false; it simply no longer grants authority to drive today.

Why this matters

If time information is lost, an old decision may be used as current authority, or today's change may be applied wrongly to the past. If continuing effects remain invisible, an expired decision may still affect the person.

Do not confuse

Historical accuracy is consistency with the facts at the decision's reference time.
Historical validity asks whether the decision was authorised and in force at that time.
Current accuracy asks whether its underlying facts remain true today.
Current validity asks whether the decision still carries legal or operational force today.
Continuing effect means that a no-longer-valid decision still produces consequences in systems.

What should you do?

Store decision time, reference time, effective start and validity end as separate fields.
Preserve the time-of-decision versions of data, rules, model, human authority and evidence.
Record active, suspended, expired, withdrawn, corrected and superseded states separately.
For a current query, keep both the current-state view and the event chain that produced it accessible.
Look for continuing effects on profiles, payments, access, recommendations and representation.
Link the new decision to the old one with an explicit relationship and effective date.
Verify the correction on affected live and downstream surfaces.

How do you audit it?

Are decision time and the reference time of the data and rule distinct?
Are the start and end conditions of validity recorded?
Is an earlier decision being used as current authority rather than as a historical record?
Has new information unfairly changed the assessment of whether the earlier decision was accurate at the time?
Can suspension, withdrawal, correction and supersession relationships be seen?
Were continuing effects of an invalid decision sought?
Is the live system selecting the current authorised decision?

Limit

Historical accuracy and legal validity are not the same, and the effect of a past decision may be determined by applicable law. A fact that appears wrong today does not by itself prove that it was wrong when the decision was made.