Conflicting finance policies
Show version status before deciding

Microsoft's Azure AI Search reference describes the $filter parameter as inclusion or exclusion criteria for search results. Its OData expressions, a language for describing those conditions, can combine comparisons and logical operators. The page includes date comparisons. These mechanisms answer whether indexed records meet a stated condition, not whether their contents govern a business decision. A later modification date could therefore help find a recent revision without establishing that it is approved, effective or relevant to the task. Ask what the date field represents and where its value came from. These are our proposed review questions, not extra authority supplied by the filter language.

Publication status introduces a separate technical surface. Microsoft's SharePoint Server planning page applies to Server 2013, 2016, 2019 and Subscription Edition. It explicitly excludes SharePoint in Microsoft 365. The page describes versioning and content approval configured in document libraries; libraries can have different settings. The two sources do not document a tested connection between that library and a search assistant. Even if a proposed implementation uses both products, whether it carries version, approval and visibility information into its index remains a question for its implementer. A source link alone does not demonstrate that connection.

The SharePoint Server page explains that, with content approval enabled, a changed document can be Pending while readers continue to see an earlier version. Under major versioning, the previous major version remains available to readers while changes await approval. After approval, the new major version becomes available and the previous one enters history. For a newly uploaded document under that configuration, readers cannot view it until approval. Major-and-minor versioning adds a draft distinction: authors can save a minor draft or submit a major version for approval. Who can see minor versions is configurable. Required check-out also affects visibility: saved edits remain private to the person editing until check-in. These are conditional publication behaviors, not a universal rule that every latest file is hidden or every major version governs. Even the no-versioning option needs context. In the content-approval scenario described, the server temporarily keeps the earlier reader-visible content while changes are pending, then does not retain that earlier version after approval. That is different from having a retrievable version history. Ask for the actual library settings rather than inferring them from a filename or version number.

We propose a compact evidence record beside the answer. For each conflicting passage, include the source link and location, document identity and version, retrieved wording, available publication or approval status, and relevant dates with their meanings. Record the task's scope and any stated applicability conditions separately. If a field is unavailable, say so rather than filling it from another date. The technical owner should explain the search conditions, indexed metadata and connection to the source version. The library owner should confirm relevant settings and the access context behind the evidence. This is not a recommendation to expose restricted drafts. Ask what authorized evidence can establish status without extending access. The responsible policy owner then decides applicability or requests clarification. Being visible to a reader is evidence about publication under particular settings, not proof of financial or legal precedence. Likewise, an approved label cannot by itself establish the task's effective date, organizational scope or exceptions.

For an evaluation, propose authorized material that lets reviewers examine a newer pending revision alongside an earlier reader-visible version. Ask whether the answer preserves the disagreement, distinguishes date meanings and shows the evidence available for each status. This is suggested evaluation work, not a test performed here or a vendor-mandated procedure. A useful handoff states what differs, what supports each version's status and what remains unknown. Where applicability cannot be established, keep the recommendation pending for its named owner rather than merging incompatible instructions. Record any subsequent decision with its basis and limits. The process figure shows this reading sequence; its final step can remain unresolved. The buying question is therefore not simply whether an assistant finds the latest policy. It is whether a reviewer can see how the candidate versions were selected, what their publication context establishes and where a human decision is still needed.