Life sciences
Trial master file: no artifact for an AI answer
The trial master file exists so anyone can judge how a trial was run from the record alone. An AI answer that shapes a trial decision is filed nowhere, and the TMF Standard Model has no artifact for it.
Cognatum Team · Nov 3, 2026 · 6 min read
A study coordinator asks an assistant one question. Which version of the protocol applies at this site today? The answer comes back in a second. She acts on it. The trial master file holds the protocol, every amendment, and the approval that made each one effective. It holds nothing about the answer.
Cognatum governs the entry
source · version · approver · permissions
What the file is actually for
The trial master file is not a filing cabinet. It is the proof. ICH good clinical practice, published in 1996, set the expectation across its regions. The file holds essential documents that, on their own and together, let someone judge how a trial was run and how sound its data is. They exist to show that the investigator, the sponsor and the monitor followed the rules.
Europe put the same idea in law. Regulation (EU) No 536/2014 says the file must hold, at all times, the essential documents that allow someone to verify how the trial was run and the quality of the data it produced. The same regulation sets a floor of 25 years for keeping it.
At all times, not at the end
Two words there do the work. At all times means the file is built while the trial runs, not assembled once an inspector books a date. A record that did not exist on the day a decision was made cannot be back-filled into evidence later.
E6(R3) turned essentiality into a test
The 2025 revision of ICH E6 changed how sponsors decide what goes in. Rather than working down a fixed list, they apply criteria. A record is essential if it documents significant trial-related activities, or if it carries the data and the relevant metadata needed to allow the appropriate evaluation of the conduct of the trial.
The drafting history is worth a look. The draft version of that criterion asked whether a record held what would be needed to reconstruct the trial. The final text replaced reconstruct with allow the appropriate evaluation of the conduct of the trial. More precise, and the same demand underneath: someone outside your organisation has to work out what happened from the file alone.
Apply the test to an answer
Now run that test on the answer the coordinator received. It shaped a decision about trial conduct. It was relied on. Under the criteria, a record of it is the kind of thing that belongs in the file. There is no record of it to file.
There is no artifact for an answer
The industry does have a shared structure for this. The TMF Standard Model, until recently the TMF Reference Model, sets out a standardised taxonomy, metadata and nomenclature for file content. It moved from the Drug Information Association to CDISC in 2022, and a survey run by the initiative that year found 96% of eligible respondents using it. Its authors are TMF managers from more than 250 organisations.
Look for the one that is missing
Look through it for the artifact that holds an AI-generated answer and there is not one. There is an artifact for the protocol. There is one for the monitoring report, the training log, the ethics approval. CDISC is careful to say the model is standard in the general sense and not a formal, enforceable standard, so a sponsor is free to add to it. Almost nobody has.
The gap is structural, not clerical
This is not an oversight by the people who built the model. Every artifact in it maps to a document somebody authored, reviewed and signed. An answer assembled on demand out of several documents has no author and no signature. It does not fit the shape of the thing being filed.
The metadata that never arrives
Metadata, in the NIST glossary, is information describing the characteristics of data: how it is structured, and what it contains. E6(R3) asks for the data and the relevant metadata together, because data on its own rarely settles anything.
An AI answer arrives as text. It does not say which version of the protocol it read. It does not say who approved that version, or when it took effect, or whether a second document in scope said something else. The MHRA's guidance on GxP data integrity sets expectations across the whole lifecycle of data for organisations working in a regulated environment. An answer with no version and no approver has no lifecycle to speak of.
Logging is not filing
Some teams assume the AI system's own records close the gap. They do not. Article 12 of the EU AI Act requires high-risk systems to allow the automatic recording of events over the life of the system, so that risks can be spotted, post-market monitoring supported, and operation watched.
Read what those logs have to contain. For remote biometric identification, the minimum is the period of each use, the reference database checked against, the input data that produced a match, and the identity of the people who verified the result. That is the one place the law asks for the source and the human by name. For every other high-risk system, a log shows that the system ran.
What a log cannot tell an inspector
A log can show that a question was asked at 09:14 and an answer came back. It cannot show that the answer came from the protocol version in force that morning. Nor can any software hand you compliance with good clinical practice or with the AI Act. A tool aligns with a framework and produces evidence for it. The obligation stays with the sponsor.
When two documents disagree
There is a harder case than a stale document. A survey of knowledge conflicts in large language models separates three kinds, among them inter-context conflict, where two retrieved sources contradict each other. In a trial that situation is ordinary: a protocol amendment, and a site-level instruction that was never withdrawn.
A retrieval system settles that quietly. It ranks, it picks, and it writes one fluent answer. Nobody is told that two answers were available and that one of them had been retired.
What filing an answer would take
The Cognatum Knowledge Loop runs eight steps across four phases, with one human gate. AI runs seven of the eight: it finds, gathers, cleans, enriches and surfaces. Approve is the step it does not run. A named person decides what is true, reworks the wording where it needs reworking, and signs.
What comes out carries what a file would need. A named approver. A date. The source it was drawn from, and where that knowledge stood at the moment the answer was given. Where two sources disagree, Cognatum puts both in front of a person and shows where the disagreement sits, rather than choosing for them.
Nothing has to move
Your documents stay where they are. Cognatum works bidirectionally with the systems that already hold them, so nothing migrates and the departmental split stays as it is. When a source changes, it tells you what has gone out of date. It notifies. It does not quietly rewrite your knowledge, and it is not current by itself.
Your company's knowledge isn't missing. It's unusable. Cognatum changes that.
Where this lands
An inspector will get to the obvious question about AI in a trial soon enough: show me what it told your team, and show me what it read. The trial master file answers the second half already. The first half is filed nowhere, which is the one thing the file was built to prevent.
Sources
- Trial master file (en.wikipedia.org)
- ICH E6(R3) Good Clinical Practice, Step 4 final guideline (database.ich.org)
- ICH GCP E6 R3: Updates to Essential Records (readyroom.net)
- Trial Master File standard and FAQ (cdisc.org)
- Metadata, Computer Security Resource Center glossary (csrc.nist.gov)
- Guidance on GxP data integrity (gov.uk)
- AI Act Article 12: Record-keeping (ai-act-service-desk.ec.europa.eu)
- Knowledge Conflicts for LLMs: A Survey (arxiv.org)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
Does an AI-generated answer belong in the trial master file?
Apply the E6(R3) essentiality criteria and the answer is often yes. If an answer documented or drove a significant trial-related activity, it is the kind of record that lets someone evaluate how the trial was conducted. There is no artifact for it in the TMF Standard Model today, and CDISC is clear that the model is not a formal, enforceable standard, so a sponsor can extend it.
What changed about essential records in ICH E6(R3)?
The guideline moved from a fixed list toward criteria. A record counts as essential if it documents significant trial-related activities, or if it carries the data and the relevant metadata needed to allow the appropriate evaluation of the conduct of the trial. The draft wording asked whether the record would let you reconstruct the trial; the final text settled on evaluation of conduct. Sponsors are not required to document the assessment itself.
Do AI system logs satisfy the record-keeping expectation?
They serve a different purpose. Article 12 of the EU AI Act requires automatic logging of events over a high-risk system's lifetime, to support risk identification, post-market monitoring and oversight of operation. A log can show that an answer was produced. It does not show which version of a source document it came from, or who approved that version, or that a second source disagreed.
Can software make us compliant with GCP or the EU AI Act?
No. No software confers compliance with either. A tool can align with a framework and produce the evidence an auditor or an inspector asks for, which is a different and more honest claim. The obligation sits with the sponsor and does not transfer to a vendor.
Do we have to move our trial master file into Cognatum?
No. Documents stay in the repositories that already hold them. Cognatum works bidirectionally with those systems and governs the knowledge above them, so the departmental split you have today is fine. When a source document changes, it notifies you about what has gone out of date rather than editing anything on its own.