Retrieving a document is not the same as knowing which one is authoritative

The current generation of enterprise AI tools is remarkably good at retrieval. Ask a question of your document estate, and Copilot, Glean, SharePoint search, or a competent RAG pipeline will return relevant passages, quickly, most of the time. This has been sold as a solution to knowledge work. In many settings it is. In a regulated setting, it isn't — and it isn't for a specific reason worth naming.

The reason is that retrieval is not authority. And a regulated filing depends on authority.

Retrieval answers the wrong question

A retrieval system answers the question "which passage of which document in my corpus is most relevant to this query?" That is a well-defined and impressive achievement. But it is not the question a signatory is asking. The signatory is asking, "of the passages that might be relevant, which one is authoritative — meaning current, non-superseded, produced by the accountable owner, and defensible if a regulator points at it?"

Those are not the same question. A retrieval system that returns three passages, one of which is from a superseded procedure, has not failed at retrieval. It has answered the retrieval question correctly. It has just answered a different question from the one that mattered.

Regulated estates are messy on purpose

A cleanly-curated document estate would make this easy. Regulated estates are not cleanly curated, and cannot be. They contain:

• Multiple live revisions of the same document, at different degrees of formal supersession.
• Procedures that have been "withdrawn" in the document management system but remain readable, cited, and physically present in operational settings.
• Standards issued by external bodies that have themselves been updated — so a document that was current when it was written may reference a version of a standard that no longer exists.
• Draft documents that were shared, worked from, and never formally issued, but which still shape the way work is done.

Any of these can be returned by a retrieval system with the same confidence as the authoritative source. Nothing about retrieval as a technology tells you which is which. This is not a criticism of the technology; it is a description of the boundary of what it does.

Evidence-grade means something specific

"Evidence-grade" is a phrase we use deliberately, and it is worth defining. An output is evidence-grade when four things are true of every claim in it.

Claim-to-source binding. The output points at the specific document, revision and passage from which the claim is drawn. Not "the source pack." Not "these documents." A specific passage.

Verbatim quotation. The passage from the source is present in the output — not paraphrased. A reviewer verifying the claim reads what the model read, and can decide whether the claim follows from it.

Authority resolution. The source has been checked against a supersession model — is it current? has it been withdrawn? does a later revision exist? — and the answer is recorded next to the claim, not left implicit.

Reproducibility. The run that produced the output can be replayed later. Same corpus, same prompt, same model, same result. If a regulator asks in a year how a claim was arrived at, the record answers it.

Any output where one of those four is missing is not evidence-grade. It might be good enough for internal use. It is not good enough for a filing.

What this means for the tool question

The most common misunderstanding in this space, in our conversations with regulated buyers, is the idea that if the retrieval quality is high enough, the authority question takes care of itself. It does not. Authority is a separate discipline. It requires an explicit model of the corpus — which document is the current owner of a given topic, which are superseded, when and by what. It requires that the retrieval system respect that model rather than treat the whole corpus as flat. It requires that the output surface the authority decision alongside the answer, so the reviewer can see and challenge it.

None of this is beyond what horizontal tools could do. It is what they do not do, because their buyer is not the signatory. Their buyer is the knowledge worker looking for a document.

For a regulated filing, that difference is the difference between a tool that saves time and a tool that survives the question. We think it is worth building for the second.