Catch stale information in an AI-assisted decision

Mark the decision's as-of time

Suppose a synthetic facilities team asks an assistant to summarize access requirements for a new site visit. It retrieves a visitor procedure from last year and writes that a printed badge request must be submitted two working days ahead. A current site notice has replaced that procedure with a same-day digital registration step for ordinary visitors, while a separate safety approval is still required for restricted areas. The old document is genuine. Its wording is not fabricated. The error is that the assistant presented an expired rule as the instruction for next Tuesday's visit.

Before drafting, write the decision date and the scope: which site, visit type, and access class? A document's publication date alone is not enough. One file may be a current parent policy with a dated local annex; another may be an archived copy kept for audit. Record who owns the rule and where the authoritative version is maintained. If there is no clear owner or effective date, do not invent one. The correct answer may be that the team must ask the facility owner before advising the visitor.

The UK Government Data Quality Framework treats timeliness and fitness for the intended use as dimensions of data quality. That principle does not make a particular visitor rule current; it tells the reviewer to test whether the evidence fits the decision date. The UK Government Digital Service describes an operational documentation review process with page owners, review dates, and actions when content is out of date. Its example is documentation maintenance, not a mandatory procedure for every organization.

Sources: [1] [2]

Build a currency card for the governing sentence

The fictional facilities team writes a card for each consequential instruction. One card reads: claim = “submit a paper badge request two days before arrival”; source = visitor procedure version 3.1; effective period = superseded by notice 4.0; decision date = next Tuesday; owner = facilities access lead; change signal = updated registration link in notice 4.0; action = remove the paper-badge instruction. A second card reads: claim = “restricted areas need separate safety approval”; source = restricted-area schedule version 2.2; status = current according to the named owner; action = keep, but specify that this applies only to the restricted-area visit.

This small artifact prevents two opposite mistakes. One is blindly using the latest file even when it is a draft with no authority. The other is using an old approved file after its effective period ended. The reviewer should look for a supersession statement, effective date, owner confirmation, and cross-reference from the current index. If sources conflict, record the conflict and hold the affected sentence rather than choosing whichever source the assistant cited first. A date stamp is evidence of when something was written, not proof of current validity.

Verify the change path, not just the link

At release, open the authoritative source location, check the version status, and ask whether the change affects any downstream artifact: visitor email, calendar invite, checklist, safety brief, or vendor instruction. Repeat the check after a material change rather than relying on a cached summary. Keep the version used for the released instruction so a later correction can explain what changed. Where a claim has a natural expiry, name a refresh owner and trigger; “review periodically” without an owner is difficult to execute.

The result is a bounded operational verification method; its effectiveness still depends on owners noticing and recording source changes. The example contains no real visitor information. In a Two Prune assessment, a participant can be asked to reconcile synthetic old and current records, and the resulting work can inform a human-interpreted evidence report. Two Prune does not maintain a customer's live procedures or make automated decisions about people.