A regulatory writer asks the enterprise model a question that should be easy: what is the approved storage condition for the drug product, and which prior submission established it? The system, grounded by retrieval-augmented generation over the company's own document library, returns a confident, well-cited answer that names a temperature range and points to a controlled storage SOP by its document number. The answer is wrong, not because the model hallucinated, but because the retrieval layer pulled the SOP that was current in 2023 and has since been superseded, and the storage condition changed in the amendment the retriever did not surface. This is the failure mode that makes grounding dangerous in a way that raw hallucination is not: the answer is sourced, it is cited, it looks exactly like the disciplined, traceable output the whole program has been teaching toward, and it is built on a deprecated source of truth. Retrieval-augmented generation is the most important architectural pattern for regulated AI, because it grounds the model in your own validated content rather than the open internet. But grounding is only as trustworthy as the source it grounds in, and this lesson is about building RAG over a submission library so that the version of record, not merely a plausible match, is what reaches the model.

What RAG Actually Does, and Why It Is the Right Pattern

Retrieval-augmented generation changes where the model's factual content comes from. Instead of relying on the patterns baked into its training weights, which know the shape of an Investigator's Brochure but not the contents of yours, a RAG system first retrieves relevant passages from a curated corpus, your IBs, your Clinical Study Reports, your prior submissions, your standard operating procedures, and places those passages into the context window alongside the question, so the model generates its answer from your documents rather than from its general training. This is the correct architecture for regulated work, because it lets the model answer questions about content it was never trained on while keeping the answer anchored to a citable source, and it is the mechanism behind every credible claim that an enterprise tool is grounded in the sponsor's own truth.

The promise is real, and it is why RAG underpins the production submission tools the industry is deploying. A grounded system can answer that the median progression-free survival in a pivotal study was a specific value and cite the CSR section it came from, can summarize the safety profile from the actual safety database rather than from generic oncology knowledge, and can surface the relevant precedent from a prior submission instead of inventing one. But the architecture introduces a new and subtle failure surface that did not exist with the raw model, and it is located not in the generation step the program has spent so long teaching writers to distrust, but in the retrieval step that precedes it. The model can reason flawlessly over the passages it was given and still be catastrophically wrong, because the passages it was given were the wrong ones. Grounding moves the risk upstream, from invention to retrieval, and a writer who has learned to verify the model's claims must now also learn to verify the model's sources.

The Source-of-Truth and Version-of-Record Problem

The deepest problem in RAG over a regulated library is that a document library is not a set of facts; it is a set of versions, and only some of them are the version of record. A submission library accretes superseded protocols, deprecated SOPs, draft CSRs that were later revised, prior-submission content that a subsequent amendment overrode, and the same numerical claim appearing in a dozen documents at different stages of its lifecycle. The retriever's job is to find passages that are semantically relevant to the question, and semantic relevance is indifferent to currency: a deprecated SOP about storage conditions is exactly as relevant, by meaning, as the current one, and is often more retrievable because the superseding amendment may phrase the change in regulatory language that matches the query less well than the original SOP's direct statement. The retriever, optimizing for relevance, can confidently surface the obsolete version precisely because it reads as the cleaner answer.

This is not a bug the retriever can solve by being smarter, because the question of which version is authoritative is not a semantic question at all; it is a governance question that lives in the document-management system's metadata, in the effective-date and status fields, in the change-control records, in the supersession links. A retriever that indexes the text but not the lifecycle status has no way to know that the passage it found best is the passage that no longer governs. The version-of-record problem is therefore the central design constraint of regulated RAG: the system must not merely retrieve relevant content, it must retrieve currently authoritative content, and that requires the retrieval layer to be governed by the same source-of-truth discipline that governs the library itself. A RAG system grounded in a library without enforced version control is grounded in a contradiction, and it will, sooner or later, cite a deprecated SOP as if it were law.

When Grounding Fails: The Anatomy of a Deprecated Retrieval

Trace the storage-condition failure from the opening in full, because its anatomy is the anatomy of every grounding failure. The writer's query is clear and the retrieval is technically successful: the system finds a passage that directly answers the question, attaches the document number, and the model faithfully generates an answer that transcribes the retrieved passage without invention. Every step the program has taught as a safeguard, source-grounding, citation, faithful transcription, has functioned correctly, and the output is still wrong, because all of those safeguards verify fidelity to the retrieved source and none of them verifies that the retrieved source is the authoritative one. The failure is invisible to the writer precisely because the answer bears every mark of a trustworthy one; the citation makes it look more reliable than an uncited answer would, which is the cruel inversion at the heart of the problem.

The failure compounds when the deprecated content propagates. The writer, reasonably trusting a cited answer, carries the obsolete storage condition into a draft Module 3 stability narrative, where it now sits with a citation that a reviewer can follow, and the citation leads to a real document, just not the governing one. A reference QC that checks whether citations resolve to existing documents passes it, because the document exists. The error is caught only if someone independently knows the storage condition changed, or if a later cross-check against the current specification surfaces the contradiction, and in a large submission assembled under deadline, that someone may not exist. The deprecated retrieval has all the propagation dynamics of a hallucinated cross-reference from the foundational lessons, with an added danger: it is harder to detect, because it is true to its source and false only to the truth.

Building Version-Aware Retrieval

The fix is to make the retrieval layer version-aware, which means the index over which the system retrieves must carry, alongside the text, the governance metadata that determines authority: the document status, the effective date, the supersession relationships, and the approval state from the document-management system. With that metadata in the index, the retrieval step can filter to currently effective, approved, non-superseded content before it ranks by semantic relevance, so that a deprecated SOP is never a candidate answer regardless of how well its text matches the query. This is the single most important design decision in regulated RAG, and it inverts the naive pipeline: governance filtering comes first, semantic ranking comes second, because an authoritative-but-less-relevant passage is always preferable to a relevant-but-deprecated one when the question is what currently governs.

Version-aware retrieval also has to handle the cases where authority is genuinely ambiguous or contested, because real libraries contain them. A claim may be current in one region's submission and superseded in another, a document may be effective but under active revision, a stability condition may differ between the original approval and a pending supplement. A well-designed system does not silently pick one; it surfaces the conflict, returning the candidate passages with their status and effective dates and flagging that the question has more than one authoritative answer depending on context, which routes the determination to the human who can resolve it. The goal is not to make the retriever decide which version governs, which is a regulatory judgment, but to ensure the retriever never hides that a version question exists. Grounding that surfaces version conflicts is trustworthy; grounding that silently resolves them is the deprecated-SOP failure waiting to happen.

Detecting Grounding Failures Before They Ship

Even a well-governed RAG system needs detection, because metadata is imperfect, supersession links break, and a document mis-tagged as current will pass governance filtering and reach the model. The first detection layer is the citation-currency check: for every source the system cites, verify not only that the document exists, the check reference QC already performs, but that it is the currently effective version, by reconciling the cited document's status against the document-management system at the moment of use. This converts the citation from a claim that a source exists into a claim that the governing source was used, which is the claim a regulated answer actually needs to make.

The second detection layer is contradiction detection across the retrieved set and against the known source of truth. When a query returns multiple passages, the system can compare them and flag where they disagree, because disagreement among retrieved sources about a factual claim is a signal that at least one is non-current or that the question spans a version boundary, exactly the condition that produced the storage-condition failure. A more rigorous layer round-trips the answer: it takes the generated claim and checks whether the current authoritative document actually contains it, so a storage condition that the model transcribed from a deprecated SOP fails the round-trip against the current specification even though it passed every fidelity check. These detection layers do not replace version-aware retrieval; they are the defense in depth that catches the residue version-aware retrieval misses, and together they make the difference between a RAG system that is usually right and one a sponsor can defend when an answer it produced is challenged.

Chunking and Retrieval Quality as a Second Failure Surface

Beyond version, a second class of grounding failure comes from how the documents were broken into retrievable pieces, and it is easy to overlook because it does not announce itself. A RAG system does not retrieve whole documents; it retrieves chunks, passages sized to fit the context window and to match queries at a useful granularity, and the chunking strategy determines what the retriever can and cannot find. Chunk too coarsely and a precise question retrieves a passage so broad that the specific answer is diluted or absent; chunk too finely and a claim is severed from the context that qualifies it, so a retrieved sentence stating an efficacy result loses the adjacent sentence that bounded it to a subgroup. In a regulated corpus, a claim retrieved without its qualifying context is a different and more dangerous artifact than the same claim in situ, because the model presents the unqualified version with the same confidence the qualified one deserved.

The retrieval-quality problem also includes the silent miss, where the authoritative passage exists in the corpus and is current but simply does not rank highly enough to enter the context window, so the model answers from a weaker but better-matching passage or, worse, declines to find an answer that is genuinely present. This is the retrieval analogue of the context-window silence from the foundational lessons: a passage not retrieved is invisible to the model, and the model does not experience the gap. A well-designed regulated RAG system therefore treats retrieval quality as a measured property, evaluating against a known set of questions with known authoritative answers to confirm that the system retrieves the right passage at the right granularity, and it tunes chunking to preserve the qualifying context that turns a fluent claim into a true one. Retrieval quality, like version-awareness, is not a one-time configuration but a property that has to be measured and maintained.

Governing the Corpus as a Validated Component

Because this is a Level 3 design, the corpus a RAG system retrieves over is not just a folder of documents; it is a validated component of a regulated workflow, and it has to be governed as one. The intended-use statement defines what the corpus is authoritative for: a corpus assembled to answer questions about a specific product's submission history is not authoritative for a different product, and a retriever that silently crosses that boundary, answering a question about product A with a passage from product B's CSR, is a grounding failure of a different kind. Corpus governance specifies what is indexed, how documents enter and leave the index as their status changes, how the index stays synchronized with the document-management system's source of truth, and how the whole arrangement is qualified to retrieve current authoritative content.

The lifecycle dimension is what most teams underestimate. A document library is not static; documents are approved, superseded, withdrawn, and revised continuously, and a RAG index that was correct at qualification drifts out of correctness the moment the library changes and the index does not. The corpus must therefore be maintained under change control, with the index updated as document statuses change, so that a SOP superseded on Tuesday is no longer a retrieval candidate on Wednesday. This is the operational heart of keeping grounding trustworthy over time, and it is why a RAG system is qualified not as a one-time event but as an ongoing arrangement whose performance is monitored, because the failure mode this lesson opened with, a deprecated SOP surfacing as if current, is precisely what an unmaintained index produces with certainty rather than chance. The model grounds in the corpus; the corpus governance ensures the corpus tells the truth; and the named owners ensure the governance holds as the library lives and changes.

Key Takeaways

• RAG grounds the model in your own validated content, but it moves the risk upstream from invention to retrieval. A grounded system can reason flawlessly over the passages it was given and still be catastrophically wrong because the passages were the wrong ones, so a writer who learned to verify the model's claims must now also verify the model's sources, and a cited answer can be less trustworthy than it looks.
• A document library is a set of versions, and the version-of-record problem is the central design constraint. Semantic relevance is indifferent to currency, so a retriever optimizing for relevance can confidently surface a deprecated SOP that reads as the cleaner answer; which version is authoritative is a governance question in the document-management metadata, not a semantic question the retriever can solve by being smarter.
• A deprecated retrieval is harder to detect than a hallucination because it is true to its source and false only to the truth. It passes source-grounding, citation, faithful transcription, and a reference QC that checks document existence, then propagates into a draft with a citation that resolves to a real but non-governing document, caught only if someone independently knows the content changed.
• Version-aware retrieval inverts the pipeline: governance filtering first, semantic ranking second. The index carries document status, effective date, and supersession relationships so deprecated content is never a candidate, and where authority is genuinely ambiguous the system surfaces the conflict for human resolution rather than silently picking one, because the retriever must never hide that a version question exists.
• The corpus is a validated component governed under change control, not a static folder. The index must stay synchronized with the document-management source of truth as documents are approved, superseded, and withdrawn, because an unmaintained index produces the deprecated-SOP failure with certainty rather than chance; citation-currency checks, cross-source contradiction detection, and answer round-tripping are the defense in depth that catches what governance filtering misses.