An Investigator's Brochure is past due for its annual update, and the safety physician who owns it is looking at a stack of new evidence: three new study Clinical Study Reports completed since the last revision, two recently published literature articles describing the drug class, and the latest Periodic Safety Update Report and Development Safety Update Report from the pharmacovigilance team. The job is to integrate all of it into the IB's safety sections so that every investigator running a trial with this compound reads an accurate, current picture of what is known about it. ICH E6(R3) defines the Investigator's Brochure, at Section 1.55 of its glossary, as a compilation of the clinical and nonclinical data on the investigational product relevant to its study in human subjects, presented in a concise, objective, balanced, and non-promotional form. That definition is the whole assignment. The IB is the document that tells an investigator whether a patient in front of them is experiencing an expected adverse reaction or a new signal, and it is the basis from which the protocol's Reference Safety Information is drawn. An update that quietly drops a known risk, misattributes a finding to the wrong source, or invents a frequency is not an editorial slip; it is a patient-safety document that has become unreliable. This lesson uses AI to integrate the new evidence into IB Section 5, Effects in Humans and Safety, and Section 7, Summary of Data and Guidance for the Investigator, with the full source traceability that the document's purpose demands.

Why the IB Update Is a Synthesis, Not a Summary

The previous two lessons dealt with documents that derive from a single dominant source: the Module 2.5.4 from a CSR, the M11 protocol from a TPP and an IB. The IB update is different in kind. It is a synthesis across multiple, heterogeneous, sometimes conflicting sources, and synthesis is the task where language models are simultaneously most useful and most dangerous. They are useful because integrating three CSRs, two papers, and two periodic safety reports into a coherent narrative is genuinely laborious, and the model is fast at finding the relevant passages and drafting connective prose. They are dangerous because synthesis invites the model to smooth over disagreements, to blend a frequency from one study with a population from another, and to assert a unified conclusion the sources do not jointly support.

Consider what integration actually requires. The three new CSRs may report the same adverse event at different frequencies because their populations and durations differ. A literature article may describe a class effect that the compound's own data neither confirm nor exclude. The PSUR or DSUR may have upgraded a risk from potential to identified, or added a new safety concern from post-market or ongoing-trial data that the CSRs predate. A faithful IB update has to hold all of this in tension: report each frequency with its source and population, distinguish compound-specific findings from class effects, and reflect the current risk characterization from the periodic reports. A model left to its own devices will instead produce a smooth, confident paragraph that picks one frequency, generalizes one finding, and reads as if the evidence were unanimous. The smoothness is the failure. The IB's value is precisely in the distinctions the smooth paragraph erases.

This reframes the verification task. For the 2.5.4 you reconciled each claim to one table cell. For the IB update you reconcile each claim to its specific source among several, and you additionally check that the synthesis did not manufacture agreement that the sources do not support. The first is a transcription check; the second is a synthesis-integrity check, and it is the harder one, because a synthesis error is not a wrong number, it is a wrong relationship between numbers, and the prose hides it well.

Staging Seven Sources With Provenance

The source set for this update is larger and more varied than for the earlier lessons, and provenance, knowing exactly which document a piece of evidence came from, is no longer a nicety but the core of the task. You load the existing IB, because the update is a revision and you need to know what the current document says in order to change it deliberately rather than rewrite it. You load the three new CSRs, each clearly labeled with its study identifier, population, and duration, because the same adverse event from different studies must remain attributable to the right one. You load the two literature articles, labeled as external literature rather than sponsor data, because a finding from the published literature carries different evidentiary weight than a finding from the sponsor's own controlled trial. And you load the latest PSUR and DSUR, which carry the current aggregate safety assessment and the official risk characterization that the IB must reflect.

The provenance labeling is what makes the synthesis checkable. When the model later writes that a particular adverse event occurred at a given frequency, the claim is only verifiable if you can tell which source it came from, and the only way to guarantee that is to label every source distinctly as you load it and to require the model to cite the specific source for every claim. A frequency attributed to the wrong study, or a class effect presented as compound-specific data, or a risk characterization that lags the latest DSUR, are all errors that become visible only when provenance is tracked and invisible when the sources blur together in the window. The source manifest here is not just a list; it is the key that lets you trace each synthesized claim back to the document and the population it belongs to.

There is a hierarchy among these sources that the workflow must respect. The periodic safety reports, the PSUR and DSUR, carry the most current and most authoritative risk characterization, because they are the formal aggregate safety assessments. The CSRs carry primary study data. The literature carries external, often less controlled, evidence. When sources conflict, the IB does not average them; it reflects the authoritative characterization and explains the constituent data. A model does not know this hierarchy unless you tell it, and an IB that treats a literature anecdote as equal in weight to the DSUR's formal risk assessment has misrepresented the state of knowledge. Loading the sources with their provenance and their relative weight is the setup that makes a faithful synthesis possible.

Drafting Section 5, Effects in Humans and Safety

Section 5 of the IB, Effects in Humans, is where the clinical safety and efficacy experience with the compound is presented, and the safety content is the part that grounds every investigator's judgment in the clinic. The update integrates the new CSR safety data and the new periodic-report assessments into the existing Section 5 narrative. The prompt that produces a defensible draft requires the model to attribute every safety finding to its specific source, to report adverse-event frequencies with the study and population they come from rather than as a single pooled number unless a proper pooled analysis exists, and to preserve the distinction between the compound's own data and class effects drawn from the literature.

The most consequential instruction is about what the model must not silently do. It must not pool frequencies across studies on its own, because pooling is a statistical operation with rules, and a naive average of percentages from differently sized studies is wrong and misleading. It must not upgrade or downgrade a risk's characterization, identified versus potential, on its own, because that characterization is a pharmacovigilance determination that lives in the periodic reports and is owned by the safety physician. It must not omit a known risk from the prior IB just because the new sources do not mention it, because absence in the new evidence is not evidence of absence, and a risk that was real last year is still real unless it has been formally retired. Each of these is a place where the model's instinct to produce a clean, unified narrative directly conflicts with the IB's obligation to be accurate, and the prompt has to forbid the smoothing explicitly.

Verification of Section 5 is a provenance reconciliation. You walk each safety claim and confirm three things: that the finding is attributed to the correct source, that the frequency matches what that source reports for that population, and that the characterization matches the authoritative periodic report rather than the model's inference. You also run a completeness check against the prior IB: every risk in the previous Section 5 either persists in the new draft or has a documented, sourced reason for its change. The risk that silently disappears is the most dangerous failure in an IB update, because no one is looking for the absence, and an investigator who no longer sees a warning may not act on a reaction the document used to flag.

Drafting Section 7, Summary of Data and Guidance for the Investigator

Section 7 of the IB, the Summary of Data and Guidance for the Investigator, is the section investigators actually act on. It distills everything in the preceding sections into a concise discussion of the known and potential risks and the practical guidance for managing them: what to monitor, what dose adjustments to consider, what reactions to expect and how serious they are. Because Section 7 is a distillation, it is the section where synthesis error is both most likely and most consequential, since a wrong synthesis here translates directly into wrong guidance at the bedside.

The drafting discipline for Section 7 is that it must be derived from Section 5 and the periodic reports, not generated independently. The model should not introduce a risk in Section 7 that is not established in Section 5, and it should not soften guidance that the safety data warrant. The prompt requires Section 7 to trace each piece of guidance to the underlying safety finding and its source, and it requires the model to flag any guidance it cannot ground in the loaded data rather than producing plausible clinical advice from its training. This last point is sharp: a model is very good at producing reasonable-sounding clinical management guidance, because such guidance is well represented in the medical literature it trained on, and that fluency is exactly the danger, because the guidance in an IB must reflect this compound's evidence, not the model's general medical knowledge.

Verification of Section 7 checks the chain from guidance back to evidence. Every monitoring recommendation, every cautionary statement, every characterization of a risk as expected or serious, must trace through Section 5 to a source. You confirm that the guidance is consistent with the authoritative risk characterization in the DSUR, that it does not understate a known risk, and that it does not introduce a recommendation the data do not support. The reference safety information that the protocol will draw from is built on this section, so an error in Section 7 propagates from the IB into every protocol's expectedness assessment, and from there into how serious adverse events are categorized for expedited reporting. Section 7 is small, and its blast radius is large.

The Cross-Document Consistency Check

An IB update is not just internally consistent; it has to be consistent with the documents that depend on it and the documents it depends on, and this cross-document check is unique to the IB's position in the safety ecosystem. The most important consistency relationship is between the IB's risk characterization and the Reference Safety Information used for expectedness assessment in pharmacovigilance. If the updated IB changes whether an adverse reaction is listed or unlisted, that change ripples into how the safety team classifies cases as expected or unexpected, which determines whether a report is an expedited fifteen-day submission. An IB update that the PV team does not reconcile against the RSI can silently change reporting obligations, and a mismatch between the IB's stated risks and the cases being reported against them is exactly the kind of inconsistency a regulatory inspector looks for.

The second relationship runs to the periodic reports themselves. The IB's Section 5 and Section 7 must reflect the same risk characterization the latest DSUR carries, because the DSUR is the formal annual safety assessment and the IB is the investigator-facing companion to it. A draft in which the IB's Section 7 describes a risk as potential while the DSUR has identified it is a contradiction between two safety documents about the same compound, and a reviewer who finds it loses confidence in both. The verification step explicitly diffs the IB's risk statements against the DSUR's, and any divergence is resolved toward the authoritative source with the safety physician's sign-off, not papered over.

This is where the AI workflow's traceability earns its keep. Because every claim in the draft was attributed to a specific source, the cross-document check is mechanical rather than archaeological: you can see, for each risk, which source the IB draft used and whether it matches the DSUR. Without that provenance, reconciling an IB against a DSUR is a manual hunt through two long documents. With it, the synthesis carries its own audit trail, and the consistency check becomes a defined, repeatable step rather than a heroic effort. The discipline that makes the AI draft safe is the same discipline that makes the downstream reconciliation tractable.

Documenting and Signing the Update Under ICH E6(R3)

The IB update closes the same way the other lessons do, with documentation that pins the run and proves the verification, but the signature here carries particular weight. The IB is a safety document owned by a named safety physician, and ICH E6(R3) treats it as part of the essential record of a trial. The AI use record captures the model and version, the system prompt, the temperature, the timestamp, and the seven sources loaded with their versions, so that the specific run that produced the integrated draft is identifiable. The verification record demonstrates that every safety claim was reconciled to its specific source, that the completeness check against the prior IB found no silently dropped risk, that Section 7 traces to Section 5, and that the cross-document check against the DSUR and the RSI was performed and reconciled.

The reason this matters more than the polish of the prose is the IB's function. This is the document an investigator consults when deciding whether a patient's reaction is expected, and the basis on which a safety case is judged serious and unexpected and therefore expediteable. A flaw in it does not stay on the page; it changes clinical decisions and reporting obligations. The named safety physician who signs the updated IB is attesting that the synthesis is faithful to the evidence, that no known risk was lost, and that the guidance reflects this compound's data rather than the model's general fluency. The AI accelerated the laborious integration; the physician owns the judgment that the integration is true. Under ICH E6(R3) and under the inspection that will eventually read the IB against the cases reported under it, that signed, traceable, completeness-checked update is what a defensible document looks like.

Note the through-line from the previous lesson. The protocol you drafted in the M11 lesson drew its Reference Safety Information from an IB, and this lesson maintains that very document as new evidence arrives. The two lessons are halves of a loop: the IB grounds the protocol, the trials the protocol authorizes generate new CSRs, and those CSRs flow back into the IB update. AI can accelerate both halves, but the loop only stays safe if every transit is traceable and every synthesis is owned by a named human. An IB update that loses a risk poisons every protocol downstream of it, which is why the completeness check and the named signature are not optional steps but the entire point.

What This Means for the Safety Physician

The safety physician facing the stack of new evidence gains a real accelerant from this workflow, but a specific one. The model is genuinely fast at the laborious part, finding the relevant safety passages across seven documents, drafting the connective narrative, and arranging the integrated text into the IB's structure. What it cannot do is decide whether a risk's characterization has changed, whether a finding is compound-specific or a class effect, whether a frequency should be pooled, or whether a risk from last year's IB still belongs, and the workflow's job is to keep those determinations with the physician while letting the model handle the mechanical integration around them.

The operating sequence carries the chapter's pattern into the IB's particular demands. Load all seven sources with distinct provenance labels and their relative evidentiary weight. Prompt for an integration in which every safety claim cites its specific source, frequencies carry their study and population, compound data are distinguished from class effects, and the model flags rather than invents any characterization or guidance. Reconcile each claim to its source, run the completeness check against the prior IB so no known risk silently disappears, confirm Section 7 traces to Section 5, and diff the IB's risk statements against the DSUR and the RSI for cross-document consistency. Document the run and the verification, and have the named safety physician sign the synthesis they have judged faithful. The next lesson moves from the investigator-facing IB to the patient-facing Informed Consent Form, where the same source must be translated not into a different document but into a different reading level, an eighth-grade ICF that still satisfies every element of 21 CFR 50.25.

Key Takeaways

• The IB update is a synthesis across heterogeneous sources, which is where models are most useful and most dangerous. Integrating three CSRs, two literature articles, and the latest PSUR and DSUR is laborious, so the model accelerates it, but synthesis invites the model to smooth over disagreements and assert a unanimity the sources do not support. The smoothness is the failure.
• Provenance is the core of the task, not a nicety. Load all seven sources with distinct labels and their relative evidentiary weight, with the PSUR and DSUR carrying the authoritative risk characterization, and require the model to cite the specific source for every claim. A frequency attributed to the wrong study or a class effect presented as compound data is visible only when provenance is tracked.
• The model must not silently pool frequencies, change a risk's characterization, or drop a prior risk. Pooling is a statistical operation with rules, identified-versus-potential is a pharmacovigilance determination owned by the safety physician, and a silently disappearing risk is the most dangerous IB failure because no one is looking for the absence. The prompt must forbid each smoothing explicitly.
• Section 7 guidance must trace through Section 5 to a source, not come from the model's general medical knowledge. Models produce fluent, reasonable-sounding clinical guidance from their training corpus, and that fluency is the danger, because IB guidance must reflect this compound's evidence. Section 7 is small and its blast radius is large, reaching every protocol's expectedness assessment.
• The cross-document check against the RSI and DSUR is unique to the IB's safety position, and traceability makes it tractable. An IB change to whether a reaction is listed ripples into expectedness assessment and fifteen-day expedited reporting, so the IB's risk statements are diffed against the DSUR and the RSI and reconciled toward the authoritative source under the named safety physician's signature, all defensible under ICH E6(R3).