01 — The Existing CSR-Automation and Clinical-Report-Automation Landscape

Research for open.csr design doc. Compiled 2026-07-25. All package versions verified against CRAN metadata (crandb.r-pkg.org) on 2026-07-25.

0. Framing: what a CSR actually is, and why automation is hard

A Clinical Study Report is structured by ICH E3, "Structure and Content of Clinical Study Reports" (adopted 1995; supplementary Q&A 2012) — 16 numbered sections plus Section 14 (in-text tables/figures) and Section 16 (appendices, including patient listings and narratives). See the EMA scientific-guideline page for ICH E3 and the ECA summary. ICH E3 has not been revised — unlike E6 (GCP), which reached R3 with EU applicability 23 July 2025 and Canadian implementation 1 April 2026 (ACRP), and unlike M11 (protocol), whose enforcement begins June 2026. The CSR is the last major regulatory document without a modern structured-content standard. That is a strategically important fact for open.csr.

The de-facto content guidance layer above ICH E3 is CORE Reference (Clarity and Openness in Reporting: E3-based), an open-access user manual published 2016 by EMWA/AMWA, with a Version 2 Terminology Table published 2019 (Hamilton et al., Research Integrity and Peer Review, 2016; V2 terminology critique of the TransCelerate CSR template, PMC6683477). CORE Reference is explicitly content guidance, not a template — which leaves an unoccupied slot for a machine-readable, testable encoding of that guidance.

Scale of the problem, from the vendor-neutral survey at IntuitionLabs: median CSR ≈ 644 pages; traditional authoring takes 3–6 months; global medical-writing market ≈ $3.6B (2021). A large share of that page count is Section 14/16 TFLs — i.e. the document is mostly generated output wrapped in prose, which is exactly the observation open.csr is built on.

1. Commercial CSR / medical-writing automation

1.1 Certara CoAuthor

The most feature-complete commercial analogue to open.csr's concept.

• Launched as CoAuthor 2023; next-generation release 17 June 2024 (press release). Descends from Synchrogenix Writer (Certara acquired Synchrogenix 2017).
• Product page claims: 275+ eCTD templates, Microsoft Word extension, structured content authoring with a "permissions-governed repository of structured content", client-specific GPT with proprietary RAG that references only allowed source data/text, traceability and version control, real-time preview, and ">90% TLF accuracy with automated analysis and summarization of tables, listings, and figures".
• Critically for us: Certara advertises that CoAuthor "directly incorporates analysis datasets, CDISC data, tables, listings, and figures within a report, providing a real-time preview." Integrates with Pinnacle 21 (Certara-owned) and Veeva RIM.
• Reported outcome: up to 40% productivity increase for drafting and QC at an unnamed biotech.
• What it does not do: the TFLs are ingested, not generated. CoAuthor summarizes an existing TFL package; it does not own the code that produced the numbers, so a "change this table" request goes back to the statistical programming team on a separate cycle.

1.2 Yseop Copilot

• Hybrid/"composite AI": symbolic reasoning + templates + LLMs (Yseop medical writing automation). Named to TIME Best Inventions 2025 (TIME).
• Document packs: CSR, Clinical Trial Narratives (CTN), Summary of Clinical Safety (2.7.4), Summary of Clinical Efficacy, Investigator's Brochure, ICF, plus preclinical PK/PD/tox.
• Integrates into Word and Veeva. Adopters named publicly since 2023 launch: Eli Lilly, Novartis, GSK, AstraZeneca. Per the IntuitionLabs survey, Yseop has been involved in 150+ trials and Novartis reportedly generated >10,000 AI-drafted reports in 2023.
• Provenance heritage is NLG (Yseop began as a rule-based NLG company, pre-LLM) — its strength is deterministic, auditable sentence generation from structured inputs; its weakness is that the structured inputs must be hand-mapped per sponsor.

1.3 TriloDocs

• Closest philosophical cousin to open.csr on the text side. Generates a first draft CSR in the TransCelerate template (adaptable on request) from protocol + amendments + SAP + the TLF package (EMWA sponsor page; GitHub org).
• Explicit design claim: "no risk of hallucinations or 'making things up'" — it extracts and re-expresses rather than free-generates, and creates in-text tables for key parameters.
• Roadmap: Module 2.7.4 Summary of Clinical Safety and Investigator's Brochure landing Spring/Summer 2026. Also does lay summaries, ICFs, patient narratives.
• Same structural limitation as CoAuthor: it consumes a finished TLF package.

1.4 Narrativa

• CSR automation with "AI Agents". Ingests TLF and ADaM datasets into a knowledge graph, then generates prose. Ships clickable source tracing — every generated text element links back to its data origin.
• Narrativa's click-to-trace is the single most relevant feature precedent for open.csr's traceability requirement, and IntuitionLabs argues bluntly that "any AI system lacking such features could be non-compliant."

1.5 Clinion CSR

• Auto-generates ICH E3-compliant CSRs from protocol, SAP, and TLFs, as a module inside a full eClinical (EDC/CTMS) suite.
• Notable feature: cross-checks values between narrative text and statistical tables, auto-flagging discrepancies — a QC concern open.csr can solve structurally (the number in the sentence is the number in the ARD) rather than by after-the-fact reconciliation.

1.6 Structured content management platforms (the "SCM" layer)

• Veeva Vault Submissions / Vault RIM — end-to-end authoring, review, approval, and assembly of submission content explicitly including clinical study reports (Veeva Submissions brief; Veeva RIM). Document-level reuse across plans/markets; 26R2 general release scheduled 7 August 2026. Veeva is the system of record, not a generator.
• Docuvera — governance-first structured content platform purpose-built for pharma; component-level create-once/reuse-everywhere, with each content block carrying metadata, source references, usage rules, and version history. This is essentially the commercial version of open.csr's Text Library, minus the code-generation link.
• RWS Tridion Docs, DITA Exchange, MadCap — DITA XML component authoring adapted to pharma; the schema acts as guardrails ("every Task must include a Safety Warning") (RWS; DITA Exchange).
• Also named in the landscape: Lexoro (RPA + NLG, 25,000-page projects), Axtria (~30% time reduction), ArisGlobal (safety/regulatory-ops focused, not CSR authoring), IQVIA (platform-level genAI).

1.7 Published evidence on LLM regulatory writing

The best public benchmark is AutoIND (Takeda + Weave), arXiv:2509.09738:

• Task: eCTD modules 2.6.2 / 2.6.4 / 2.6.6 nonclinical written summaries.
• Time: ~100 h → 3.7 h for 18,870 pages / 61 reports (IND-1); → 2.6 h for 11,425 pages / 58 reports (IND-2) — ~97% faster first draft.
• Quality (blinded scorer, 7 categories, 0–3 scale): 69.6% and 77.9%; no critical regulatory errors; deficiencies concentrated in emphasis, conciseness, clarity.
• Conclusion: "expert regulatory writers remain essential to mature outputs to submission-ready quality."

Read for open.csr: the LLM's measured weakness is judgment and emphasis, not fact assembly. That argues for an architecture where facts come from code/ARDs (deterministic) and the LLM is confined to assembling and phrasing curated blocks — which is exactly the Text Library design.

2. Open-source efforts adjacent to CSR building

2.1 Roche/Genentech NEST (insightsengineering) — the TLG production layer

The most mature open-source TLG stack. From the NEST landing page:

Also: the TLG Catalog — a curated, executable catalog of tables, listings and graphs, open-sourced 2023; and the Biomarker Catalog. NEST supported the first R-based submission to FDA and won the Roche PD Breakthrough Award (R Consortium).

Strength: industrial-grade, submission-proven table engine with a public catalog of standard displays. Gap: NEST stops at the output object. There is no document assembly, no ICH E3 structure, no prose, no report-level traceability manifest. autoslider.core proves the "assemble many outputs into a deliverable" pattern works — but the deliverable is a PowerPoint deck for internal review, not a CSR.

2.2 The ARD/ARS layer — the newest and most relevant substrate

• CDISC Analysis Results Standard (ARS) v1.0, released April 2024 (CDISC; model + user guide; GitHub). ARS closes the long-standing gap that ADaM standardized analysis data but nothing standardized the link from a table back to that data. PHUSE US 2025 paper DS09 ("Selecting GeARS") catalogs implementation strategies from macro-library integration to fully automated TFL generation.
• R implementations: cards 0.8.1, cardx 0.3.4, gtsummary 2.5.1 (2026-05-30) — gtsummary was refactored to use cards/cardx as its backend, so you can extract an ARD from a table and build a table from an ARD (ARD-first tables vignette); tfrmt 0.4.0 (GSK) — a display-metadata language automating ARD → table; siera 0.5.6 (Clymb Clinical) — generates analysis programs from ARS metadata.
• This layer is ~2 years old and still consolidating. It is the right substrate to build on, and it is early enough that a reference implementation carries outsized influence.

2.3 pharmaverse

• pharmaverse.org — the multi-company open-source clinical reporting ecosystem (originating from an R/Pharma 2021 talk). The TLG page of the e2e clinical guide is the canonical recommendation set: rtables/chevron (Roche), Tplyr 1.3.3 (Atorus), gtsummary, tfrmt (GSK), tidytlg 0.12.0, rlistings, ggsurvfit, cards/cardx, siera, plus docorator (GSK — headers/footers framing across file types) and gridify (headers/footers on tables and figures).
• pharmaverse examples — a living collection of executable end-to-end examples from raw data → SDTM → ADaM → TLG, using pharmaverseraw / pharmaversesdtm / pharmaverseadam 1.3.0 (2026-02-20). Includes "automated reporting of outputs via slides or docs."
• Gap: examples are per-output vignettes. There is no pharmaverse example that assembles an ICH E3-structured document.

2.4 R Consortium Submissions Working Group — the regulatory-credibility layer

From the 2026 plans post and WG site:

Also: FDA expanded eCTD file-format support for R packages through this collaboration. Supporting tooling: pkglite (bundle R packages into a submittable text file), r2rtf 1.3.1 (Merck), logrx 0.4.0 (execution logs for traceability).

Note the shape: every pilot submits datasets, programs, and apps. None submits a report document. Pilots 6 and 7 (AI + simulated data, both launched six months ago) are the working group publicly signalling exactly the territory open.csr occupies.

2.5 r4csr — the closest thing to a public CSR playbook

"R for Clinical Study Reports and Submission" — Zhang, Xiao, Anderson, Zhu (Merck). Organized around ICH E3 requirements, covering TLF delivery (disposition, populations, baseline, efficacy, AE), project management/setup, and eCTD submission. Built on r2rtf (merck.github.io/r2rtf) + pkglite. Limitation stated by the book itself: it emphasizes creating TLFs and submission documentation, not generating a complete CSR end-to-end, and is labelled a work-in-progress draft. It is a book, not a runnable library or app.

2.6 TransCelerate — the template/content-reuse layer

From the Clinical Content & Reuse assets page:

Two facts matter enormously for open.csr: (1) there is a free, ICH E3-mapped CSR Word template with reusable content — the raw material for a Text Library; (2) the eTemplate program was frozen after the 2024 release and the eCSR add-in source is not discoverable as a public GitHub repo (TransCelerate's GitHub org hosts 8 repos, all DDF/CDISC-rules-engine, none eCSR). The industry's structured-CSR effort has stalled at a Word macro.

Related TransCelerate work: Digital Data Flow (DDF) with CDISC, built on the Unified Study Definitions Model (USDM) (TransCelerate; CDISC DDF); a DDF Solution Showcase was co-hosted with CDISC in April 2026. DDF digitizes the protocol end of the lifecycle; nothing equivalent exists at the report end.

2.7 OpenStudyBuilder (Novo Nordisk) — the model to emulate organizationally

openstudybuilder.com / DDF context / GitHub.

• Open-source, metadata-driven study definition tool; ~300 production users at Novo Nordisk, external contributions from Boehringer Ingelheim.
• Grounded in CDISC 360i, TransCelerate DDF with USDM 4.0, and ICH M11. DDF API v3 endpoint implements USDM 2.7.1 (from v0.9); USDM v3 expected v0.10/0.11. Exports USDM-compliant JSON or M11-formatted HTML protocol documents.
• This is the single best precedent for open.csr: a sponsor-originated open-source project that owns one document type end-to-end, anchors on a public standard, and produces a real document artifact. OpenStudyBuilder did it for the protocol; the CSR slot is empty.

2.8 Other adjacent open source

• clinDataReview (UCB) — validated open-source interactive medical/safety monitoring reports; notable for its explicit validation + CI + traceability framing (PMC11271019).
• grstat — academic oncology CTU standardized statistical reporting toolbox with reporting templates in beta (arXiv:2601.13755).
• Quarto (quarto R pkg 1.5.1) / officer 0.7.6 / officedown 0.4.1 / flextable 0.10.0 — the generic document-assembly substrate. Widely used for reproducible reports; no published CSR-specific Quarto pipeline exists as a maintained open-source project. Note officedown is the least actively maintained of the set (last release 2025-05-20), which is a dependency-risk consideration.
• Explicit search finding: as of July 2026, searching GitHub and the web for an open-source project that calls itself a CSR builder returns nothing — only vendor pages and generic Quarto/publishing tooling. There is no incumbent to displace.

3. Gap-analysis table

Legend: ●●● strong / ●●○ partial / ●○○ weak / ○○○ absent.

Reading the table: the two halves never meet

The landscape splits cleanly and the split is the whole opportunity:

• The right half (open source) owns number generation — code-first, versioned, tested, FDA-exercised — and stops at the output object. Nothing open source assembles a CSR.
• The left half (commercial) owns document assembly and prose — ICH E3 templates, content reuse, LLM drafting — and treats the TFL package as an opaque, already-final input. Nothing commercial owns the code that made the numbers.

The seam between them is exactly where CSR pain lives: a reviewer says "split this table by baseline severity," and today that becomes a ticket to a statistical programmer, a new TFL package, a re-import into the authoring tool, and a re-reconciliation of every number quoted in prose. Clinion's headline feature — cross-checking numbers between text and tables — is a symptom of that seam. Narrativa's click-to-trace is a patch over it.

open.csr's genuine gap: it is the only design in which the change request, the code edit, the regenerated number, and the sentence that quotes it are the same versioned transaction. Nobody occupies this. Three supporting gaps make it defensible:

1. No open-source ICH E3 document model. TransCelerate's eCSR is Word + a frozen add-in; CORE Reference is prose guidance; ICH E3 itself has never been revised. A machine-readable, testable E3 section model with slots for TFLs and text blocks does not exist in public.
2. No public reference implementation of ARS v1.0 → display → document. ARS is 2 years old; siera, cards, tfrmt each do a segment; nobody has published the chain end to end with evidence.
3. No published test-evidence framework for a report-generation system. The R Consortium pilots produce ADRGs; NEST produces unit tests; commercial vendors produce vendor validation packages behind an NDA. A public requirements-matrix + per-component evidence site (the safety.viz pattern) applied to a CSR builder would be the first of its kind — and is the credibility mechanism that lets an open-source project be taken seriously against validated commercial tools.

Positioning summary (one line each)

• vs CoAuthor/Yseop/TriloDocs/Narrativa: they draft prose around your TFLs; open.csr owns the TFLs, so "change the table" is a code edit, not a new vendor cycle.
• vs NEST/chevron/pharmaverse: they stop at the display; open.csr adds the E3 document model, the text layer, and the assembly.
• vs R Consortium pilots: they prove R-generated data and programs pass FDA review; open.csr extends the same evidence discipline to the document.
• vs TransCelerate CC&R: it provides the ICH E3-mapped template as a frozen Word asset; open.csr makes it live, executable, and diffable.
• vs OpenStudyBuilder: it did this for the protocol via USDM/M11; open.csr does it for the CSR via ARS/E3.

4. Concrete assets open.csr can adopt or must not duplicate

Sources

• Certara CoAuthor product page · Next-gen CoAuthor press release (17 Jun 2024)
• Yseop Medical Writing Automation · TIME Best Inventions 2025: Yseop Copilot
• TriloDocs (EMWA sponsor profile) · TriloDocs GitHub
• Narrativa CSR automation · Clinion CSR automation
• Veeva Vault Submissions brief · Veeva RIM · Docuvera
• IntuitionLabs: CSR Automation — AI Opportunities & Risks
• AutoIND / Human-AI collaboration in regulatory writing (arXiv:2509.09738)
• NEST · TLG Catalog · chevron on CRAN · chevron GitHub · autoslider.core on CRAN
• R Consortium: NEST & admiral clinical reporting
• R Submissions WG: 2026 plans & 2025 success · Submissions WG site · submissions-wg GitHub · Pilot 3 repo
• CDISC Analysis Results Standard v1.0 · ARS model site · ARS GitHub · PHUSE US 2025 DS09 "Selecting GeARS"
• pharmaverse · e2e clinical TLG guidance · pharmaverse examples · gtsummary ARD-first tables
• r4csr — R for Clinical Study Reports and Submission · r2rtf
• TransCelerate Clinical Content & Reuse assets · TransCelerate DDF · CDISC DDF · TransCelerate GitHub
• OpenStudyBuilder · OpenStudyBuilder DDF context · OpenStudyBuilder GitHub
• ICH E3 (EMA) · ICH E3 (ECA Academy)
• CORE Reference development paper (Res Integr Peer Rev, 2016) · CORE Reference V2 terminology / TransCelerate CSR template critique
• clinDataReview (PMC11271019) · grstat (arXiv:2601.13755)
• ICH E6(R3) impacts (ACRP, Feb 2026)