Clinical trial protocols and amendment histories for protocol authoring AI

What public protocol sources already cover

Public sources give you breadth of registry metadata and a thin, final-version slice of full protocols. The ClinicalTrials.gov download is updated daily and covers purpose, recruiting status, conditions, design, phase and eligibility criteria summaries [3]. Benchmarks such as TrialBench convert those XML records into 23 AI-ready trial-design datasets, using trials registered before February 2024 [2]. Research catalogs such as ELIXIR Luxembourg publish protocol metadata profiles, not protocol bodies [4].

Full documents are public only in a narrower case. Under the FDAAA final rule (42 CFR Part 11), results submissions for many applicable clinical trials must include the protocol and a separate SAP, reflecting approved amendments, and those PDFs are posted publicly. Verify the current rule text on eCFR before relying on it. Responsible parties may redact personal details and confidential commercial information, so posted documents are often incomplete.

The practical consequence for model builders:

• You get one consolidated, often redacted, final version, not v1.0 through v5.0.
• The summary-of-changes table that explains why each amendment happened is usually absent.
• Phase 1 trials, non-US studies outside the rule, and terminated programs are thin or missing.
• Text extracted from PDF/A uploads loses section structure unless you rebuild it.

The 2024 protocol-authoring study built its corpus from public ClinicalTrials.gov protocols for US-sponsor Type II diabetes trials, enriched with commercial trial metadata [1]. That design shows both what public data enables and its ceiling: one therapeutic area, final versions only.

Why amendment rationale is the label worth paying for

Amendment rationale turns a document archive into supervised training data for amendment-risk and authoring models. Each amendment pairs a before-state (the prior protocol section), an after-state (the revised text) and a reason (for example, enrollment shortfall, safety signal, regulator request, operational infeasibility or a dosing change). With those triples, you can train a model to flag sections in a draft that historically drew amendments, or to propose language that avoids them.

Without rationale, you can still diff versions, but you cannot separate avoidable amendments (an inclusion criterion that excluded most of the target population) from unavoidable ones (a new safety finding). That distinction is what a sponsor's clinical operations team wants a model to learn. Ask for amendment reasons in the sponsor's own categories, plus any free-text justification sent to IRBs or ethics committees, and map them to a shared taxonomy afterward.

Useful companion records include the IRB or ethics committee approval date per version, the countries where each version applied, and enrollment counts at the time of each amendment. Even coarse enrollment milestones let you test whether a model predicts amendments before they happen rather than describing them after the fact.

ICH M11 and structured protocol data

ICH M11 gives the industry a common structured protocol template, which makes section-level alignment across sponsors much easier. The M11 package (guideline, clinical implementation template and technical specification) was reportedly adopted at ICH Step 4 in late 2025 and is moving through regional implementation; confirm current status on the ICH site. As of October 2026, sponsor adoption is uneven, and most historical protocols predate it.

For buyers, M11 matters in two ways. First, it supplies a target schema: map legacy protocols into M11 sections (objectives and endpoints, trial design, population, interventions, assessments, statistical considerations) so that authoring models learn a consistent structure. Second, any M11-native protocols you license are already segmented, which cuts extraction cost. Ask suppliers which template each document followed and keep the original headings next to your mapping, because forced mappings hide authoring conventions you may want the model to learn.

Data layout for a protocol and amendment corpus

A clean delivery separates documents, versions, sections and amendment events so each can be joined without re-parsing PDFs. Store source files as delivered (DOCX or PDF), extracted text as JSON Lines with one section per record, and amendment events as a separate table keyed to version pairs. Keep eligibility criteria as individual criterion records with inclusion or exclusion flags, because eligibility models train on criterion-level text, not paragraphs.

Illustrative example: invented to show structure; it does not describe an available dataset.

Two failure modes recur. Version labels collide when country-specific amendments run in parallel with global ones, so require an explicit parent version. Tracked-changes DOCX files sometimes carry reviewer names and comments that the supplier forgot to strip, which is both a privacy and a confidentiality problem.

Rights, privacy and who can license protocols

Protocol licensing has to come from the sponsor that owns the study documents, not from a CRO, site or vendor that merely holds a copy. CRO master services agreements typically assign work product to the sponsor, and site agreements restrict use of the protocol to running the trial. Ask any supplier to show the chain: who authored each version, who owns it, and whether any co-development or licensing partner holds rights.

Protocols contain little patient data, but they are not identifier-free. Expect principal investigator names, site addresses, medical monitor contacts and sponsor staff on signature pages and contact lists. Narratives in amendments occasionally describe individual safety events in enough detail to matter. If patient-level health information from a HIPAA covered entity or business associate is in scope, de-identify it through Safe Harbor or Expert Determination [5]; see our guide to Safe Harbor versus Expert Determination for training data.

Confidentiality is the larger issue. Unpublished protocols for active programs reveal dose selection, endpoints and strategy, so sponsors will usually release closed, completed or discontinued programs first, and may require redaction of compound codes.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Buyer request checklist for protocol corpora

A precise request is easier to match and rights-review than "protocols, any phase". Use this checklist when writing a specification, and pair it with your acceptance criteria for licensed training data.

Illustrative example: invented to show structure; it does not describe an available dataset.

• Scope: phases, therapeutic areas, years, program status (completed, terminated, active excluded).
• Version depth: all approved versions per protocol, minimum number of protocols with at least one amendment.
• Amendment labels: summary-of-changes tables, reason categories, IRB justification text.
• Companions: SAP versions, informed consent template versions, eligibility criteria as discrete items.
• Format: original DOCX or PDF plus section-level JSON Lines; template used (legacy or M11).
• Redaction: investigator and staff names, sites, compound codes; method recorded.
• Allowed uses: SFT, extraction, RAG, evaluation; whether outputs may be used in sponsor-facing products.
• Exclusions: anything already public on ClinicalTrials.gov, if you only want the private delta.

Related documents have their own guides: trial master file documents for TMF classification and clinical study reports for medical writing AI. For version-linking patterns borrowed from contracts, see contract families linked to amendments.

How SourceX sources protocol and amendment data

SourceX sources operational datasets from US companies on request and manages the commercial process, including the license and ongoing purchases. Nothing is held in stock and a request does not guarantee a match. You describe the data you need, and SourceX looks for US businesses that hold it; each release is approved by the supplying company.

Every dataset is rights-reviewed for ownership and consents and delivered under a license that defines the records, uses, term and delivery. Personal details are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. Health records require HIPAA de-identification. See the industry data hub, healthcare buyers and enterprise document datasets, or submit a protocol corpus request.

Request a clinical trial protocol dataset

If you need protocol versions, amendment rationale and SAPs beyond what registries publish, describe the scope, version depth and allowed uses you need. SourceX looks for US companies that hold matching data, reviews rights and agrees terms in a license before anything is delivered. Start your request on the buyers page.