2
CSRs final
3
Not yet started
12w
Lock to final
16
ICH E3 sections
Contents
1. The Definitive Account, Not a Summary
A clinical study report is the complete account of one trial: what was planned, what was done, what happened, and what the sponsor concludes from it. It is written to ICH E3, it runs to thousands of pages with its appendices, and it is the document a reviewer reads when they want to know what the submission's summaries are based on.
| It is not | Because |
|---|---|
| A summary of the protocol | The protocol says what was intended. The CSR says what occurred, including every deviation from that intention. |
| A presentation of results | ⚠ It reports the pre-specified analyses in the pre-specified order, whatever they show. A report that leads with its best finding has stopped being a CSR. |
| Written after the fact | Its analyses were fixed in the statistical analysis plan, signed before the database was locked. The CSR executes a plan; it does not design one. |
| The submission | The NDA contains the CSRs and also contains new analyses across them — see §5. |
The single discipline that makes a CSR trustworthy: it reports what was pre-specified, in the
order it was pre-specified, including the parts that did not work.
Every element of that is checkable by a reviewer. The objectives are quoted from the protocol rather than paraphrased. The analysis populations were assigned during blind data review, before anyone was unblinded. The testing hierarchy is followed to the point of first failure, and everything below that point is presented as descriptive rather than as a claim.
A CSR is not persuasive because it argues well. It is persuasive because its structure makes selective reporting visible.
Every element of that is checkable by a reviewer. The objectives are quoted from the protocol rather than paraphrased. The analysis populations were assigned during blind data review, before anyone was unblinded. The testing hierarchy is followed to the point of first failure, and everything below that point is presented as descriptive rather than as a claim.
A CSR is not persuasive because it argues well. It is persuasive because its structure makes selective reporting visible.
2. The Structure
| ICH E3 | Section | What it contains | The part that matters |
|---|---|---|---|
| 1-5 | Title page through ethics and administrative structure | Synopsis, ethics committee record, investigators and administrative structure | The synopsis is the most-read part of the entire report and is often written last. |
| 6 | Investigators and study administrative structure | Every investigator, site and vendor with a role in conduct | Reconciles to the site list and the transfer of obligations. |
| 7-8 | Introduction and study objectives | Rationale and the objectives exactly as the protocol stated them | ⚠ Objectives are quoted from the protocol, not restated. A CSR that paraphrases its own objectives has begun to drift from what was pre-specified. |
| 9 | Investigational plan | Design, randomization, blinding, treatments, and every protocol amendment with its effect | The amendment sub-section is where a reviewer looks first for changes made after the data started arriving. |
| 10 | Study participants | Disposition, protocol deviations, and the analysis populations | ⚠ Populations are assigned during blind data review, BEFORE unblinding. This section is where that sequence is evidenced. |
| 11 | Efficacy evaluation | The pre-specified analyses in the pre-specified order, following the testing hierarchy | Testing stops at the first failure; everything below becomes descriptive. |
| 12 | Safety evaluation | Exposure, adverse events, deaths and serious events, laboratory findings | Reconciles exactly to the safety database. Two systems, one truth. |
| 13 | Discussion and overall conclusions | The sponsor's interpretation, clearly labeled as interpretation | ⚠ The only section where the sponsor argues rather than reports, and the one a reviewer reads most skeptically. |
| 14 | Tables, figures and graphs referred to but not in the text | The full TLF package | Produced from the analysis datasets, traceable back through ADaM and SDTM to source. |
| 15 | Reference list | Literature cited | Short, and routinely wrong if left to the end. |
| 16 | Appendices | Protocol and amendments, sample CRF, IRB list, investigator CVs, randomization scheme, audit certificates, participant data listings | ⚠ Larger than the report itself. Appendix 16.1 is effectively a second Trial Master File. |
Section 13 is the only place in the document where the sponsor argues, and it is labeled as
such.
Everything from section 7 to section 12 is reporting: this was the plan, this is what the analysis produced. Section 13 is where the sponsor says what it believes that means — and separating the two is what allows a reviewer to accept the first while disagreeing with the second.
A CSR that threads interpretation through its results sections has made itself harder to accept, not easier, because a reviewer who distrusts one conclusion now has to re-examine every number it was mixed into.
Everything from section 7 to section 12 is reporting: this was the plan, this is what the analysis produced. Section 13 is where the sponsor says what it believes that means — and separating the two is what allows a reviewer to accept the first while disagreeing with the second.
A CSR that threads interpretation through its results sections has made itself harder to accept, not easier, because a reviewer who distrusts one conclusion now has to re-examine every number it was mixed into.
Appendix 16 is larger than the report. It carries the protocol and every amendment, the sample case report form, the investigator list, the randomization scheme, audit certificates and participant data listings — effectively a second Trial Master File, assembled for one study. It is also where a reviewer goes when a number in section 11 does not look right.
3. What Exists Today
Two reports are final. Three do not exist and cannot.
| Study | Status at 15 Oct 2026 | Final | Note |
|---|---|---|---|
| Phase 1 SAD/MAD | Final | 2024-05 | Signed by the medical monitor and the statistician. |
| Phase 2 dose-ranging | Final | 2026-02 | The report the End-of-Phase-2 meeting and the Gate 4 case were built on. |
| Pivotal 301 | Not started | planned 2028-08 | ⚠ Cannot begin before database lock. Draft to final is budgeted at 12 weeks. |
| Pivotal 302 | Not started | planned 2028-08 | As above, produced in parallel. |
| CV sub-study | Not started | planned 2028-09 | Nested in Pivotal 301; reported separately because CR-02 scoped it separately. |
No pivotal report can begin before its own database lock, which is
21 months away.
This is not a resourcing choice or a sequencing preference — it is arithmetic. The CSR reports analyses run on a locked database, and a database that is still accruing data has no analyses to report. The fact base asserts it: a report marked final before its lock date fails the build.
The Phase 2 report, by contrast, is final and load-bearing. It is the document the End-of-Phase-2 meeting was built on and the evidence the Gate 4 business case rests on — which is why the 5.1% gastrointestinal discontinuation figure in it has been quoted in nearly every artifact since.
This is not a resourcing choice or a sequencing preference — it is arithmetic. The CSR reports analyses run on a locked database, and a database that is still accruing data has no analyses to report. The fact base asserts it: a report marked final before its lock date fails the build.
The Phase 2 report, by contrast, is final and load-bearing. It is the document the End-of-Phase-2 meeting was built on and the evidence the Gate 4 business case rests on — which is why the 5.1% gastrointestinal discontinuation figure in it has been quoted in nearly every artifact since.
Lock to final CSR is budgeted at 12 weeks, and that is already aggressive for a report of this size. It sits directly on the critical path: database lock July 2028, submission October 2028, and almost nothing between them that can be compressed. The two pivotal reports are produced in parallel by separate teams for exactly that reason.
4. Where the Numbers Come From
Every figure in a CSR is traceable back through four transformations to a mark somebody made on paper at a site.
| Stage | What it is | Who can re-derive it |
|---|---|---|
| Source | The clinical record at the site — the nurse's note, the scale reading, the laboratory report | The site, and a monitor comparing source to eCRF |
| eCRF | Transcribed into the electronic case report form | Data management, through the audit trail |
| SDTM | Standardized study data tabulation model — the raw data in a regulator-defined structure | ⚠ The reviewer, independently |
| ADaM | Analysis datasets, one row per analysis-ready observation, with derivations documented | ⚠ The reviewer, independently |
| TLFs | The tables, listings and figures in sections 11, 12 and 14 | Anyone with the ADaM datasets and the SAP |
SDTM and ADaM exist so that a reviewer can re-run the sponsor's analysis without the sponsor's
help.
That is the entire point of standardized submission data, and it changes what a CSR is for. The report is not the evidence — the datasets are. The report is the sponsor's account of what the datasets show, offered alongside the datasets so that the account can be checked.
Which means the most consequential thing about a CSR is that everything in it can be falsified. A number that cannot be re-derived from the analysis datasets is not a finding; it is an assertion, and it will be treated as one.
That is the entire point of standardized submission data, and it changes what a CSR is for. The report is not the evidence — the datasets are. The report is the sponsor's account of what the datasets show, offered alongside the datasets so that the account can be checked.
Which means the most consequential thing about a CSR is that everything in it can be falsified. A number that cannot be re-derived from the analysis datasets is not a finding; it is an assertion, and it will be treated as one.
The define.xml that accompanies the datasets is not documentation about
them — it is how they are read. It carries the derivation for every variable, and without it
the datasets are a large amount of data nobody can interpret.
5. CSRs Are Not the Integrated Summaries
| Document | How many | What it actually is |
|---|---|---|
| Clinical study report | One per study | Reports what happened in that study, against that study's protocol and SAP. |
| Integrated Summary of Safety | One per submission | ⚠ A NEW ANALYSIS pooling participants across studies. Not a compilation of the CSRs — it re-analyzes pooled data and can surface findings no individual CSR contains. |
| Integrated Summary of Efficacy | One per submission | As above, for efficacy. Both are written from the analysis datasets, not from the reports. |
| Clinical Overview (Module 2.5) | One per submission | The sponsor's argument, ~30 pages. The part of the dossier most likely to be read in full. |
The Integrated Summary of Safety is a new analysis, not a compilation of the reports.
It pools participants across every study in the program and re-analyzes them together. That produces things no individual CSR can contain: a rare event appearing three times across three trials, none of which was notable in isolation; an exposure-response relationship visible only at the pooled sample size; a subgroup finding underpowered in every study and adequately powered across all of them.
This is the most common misunderstanding about NDA content, and it has a schedule consequence. Teams that treat the integrated summaries as an assembly task budget days for them. They are original statistical work, they are written from the analysis datasets rather than from the reports, and they take months.
It pools participants across every study in the program and re-analyzes them together. That produces things no individual CSR can contain: a rare event appearing three times across three trials, none of which was notable in isolation; an exposure-response relationship visible only at the pooled sample size; a subgroup finding underpowered in every study and adequately powered across all of them.
This is the most common misunderstanding about NDA content, and it has a schedule consequence. Teams that treat the integrated summaries as an assembly task budget days for them. They are original statistical work, they are written from the analysis datasets rather than from the reports, and they take months.
It also explains a sequencing constraint. The integrated summaries cannot be produced before the individual studies' analysis datasets are final — which is after lock, in parallel with the CSRs rather than after them. The submission is not assembled sequentially from finished reports; several of its most demanding components are being written simultaneously from the same datasets.
6. What Makes a CSR Fail Review
Not wrong results — those are a program problem. These are reporting failures, and they are avoidable.
| Failure | What it looks like | Why it is fatal to credibility |
|---|---|---|
| Post hoc analysis presented as pre-specified | An analysis that does not appear in the signed SAP, reported in section 11 alongside those that do | ⚠ The reviewer compares the CSR against the SAP. Anything extra is immediately visible, and it casts doubt on every analysis that was legitimate. |
| Populations assigned after unblinding | Analysis population definitions dated after the unblinding authorization | Two dates in the file settle it. The result becomes unfalsifiable, and unfalsifiable is not a synonym for favorable. |
| Deviations under-reported | Section 10 listing fewer deviations than the monitoring reports in the TMF contain | Reconciliation against Zone 5 is routine. A gap is read as concealment even when it was carelessness. |
| Safety numbers that do not reconcile | Section 12 disagreeing with the safety database | ⚠ The two are maintained separately by design. A discrepancy means one of them is wrong and the sponsor did not notice. |
| Conclusions unsupported by the hierarchy | A claim in section 13 for an endpoint below the first failure in the testing sequence | Testing stops at first failure. A claim below that point is not a strong claim — it is not a claim. |
Every failure on that list is a failure of sequence or of reconciliation, not of science.
Each is detectable by comparing the report against a document that already exists — the SAP, the unblinding record, the monitoring reports, the safety database. None requires the reviewer to re-run a single analysis.
Which is the argument for why the lock sequence and the trial master file are program-management concerns rather than clinical ones. The credibility of a $243 million evidence package rests on a set of dates being in the right order and a set of numbers agreeing with each other, and both of those are things a program manager can actually control.
Each is detectable by comparing the report against a document that already exists — the SAP, the unblinding record, the monitoring reports, the safety database. None requires the reviewer to re-run a single analysis.
Which is the argument for why the lock sequence and the trial master file are program-management concerns rather than clinical ones. The credibility of a $243 million evidence package rests on a set of dates being in the right order and a set of numbers agreeing with each other, and both of those are things a program manager can actually control.