Cognatum

Life sciences

Design history files have no page for an answer

A design history file records the date, the revision, the participants and the result behind every design decision. An AI answer that restates the same decision records none of those. It joins the chain nowhere.

Cognatum Team · Nov 17, 2026 · 6 min read

An engineer needs to know why a part is made of one material and not the cheaper one. The reason was settled years ago, by people who have since left.

Served to one approved entry
AI assistants & agents
Proposal tools
Internal search & chat
Customer portals
Compliance & audit

Cognatum governs the entry

source · version · approver · permissions

There is a file that answers this. It holds the requirement, the review that approved it, the test that proved it, and the name of everyone who was in the room.

She does not open the file. She asks the company assistant, and a clean paragraph comes back.

The paragraph is right. It is also the only part of this story that nobody writes down.

A file for reasons, not drawings

A design history file is a compilation of documentation describing the design history of a finished medical device. It is not the recipe for building one. It is the record of how the design came to be what it is.

The requirement arrived with the Safe Medical Devices Act of 1990. Before it, inspectors could examine production and quality control records, which showed how a device was made but not how it was decided.

The change was not cosmetic. It moved the regulator's attention upstream, from the factory floor to the drawing board, and it did so by insisting the thinking be written down.

Why the file was invented

The agency explained its reasoning when the rule was finalized. Looking at six years of voluntary recall actions, it found the quality problems behind them were attributed to errors or deficiencies designed into the devices, which adequate design controls may have prevented.

So the file exists to keep a decision inspectable long after the people who made it have gone. That is a knowledge problem written into law.

What a single entry carries

Each design review is documented with the date it happened, the people who took part, the design revision under review, and the result. Not a conclusion on its own. A conclusion with its circumstances attached.

Verification follows the same pattern. The record names the date, the participants, the revision verified, the method used and the outcome.

Validation adds the identification of the design, the methods, the date and the individuals who performed it. Three different activities, one habit.

Four things, repeated everywhere

What was decided. Which revision it was decided against. Who decided it. When. A reader arriving a decade later can rebuild the decision out of those four, and the file is built so that they can.

There is a fifth rule underneath them. The design output has to be directly traceable to the design input. The file is a chain, not a drawer.

The file changed its name this year

Since February 2026 the American device regulation incorporates the international quality management standard by reference, and the old design controls section now sits elsewhere in the rule. A technical amendment published in December 2025 updated the cross references across the classification regulations.

The design history file became the design and development file. The standard asks for the same thing under another name: a file documenting the design history of a device, holding the records of changes to its design and development.

The rename closes nothing

Both versions require the file. Neither version has a place in it for an answer that was derived from the file and handed to somebody who needed it quickly.

Set the paragraph beside the file

The assistant's reply restated a design decision accurately. Put it next to the entry it came from and four things are missing.

No revision

The file entry names the revision it reviewed. The answer names nothing, so there is no way to establish later whether it read the input in force or one that had already been superseded.

No participants

A review lists its people. An answer lists nobody. If the reasoning turns out to be wrong, there is no judgment to examine, because no judgment was exercised.

No date against a source

A record is dated against the document as it stood. An answer carries the moment it was produced, which is a different fact, and it points at nothing anyone can open.

No trace

Traceability runs from input to output and back again. The answer joins that chain at no point, which means it can be followed neither forward nor backward.

Design changes are where this bites

Design changes are identified and documented so that when an input moves, everything downstream of it can be found and verified again. The register exists to answer one question: what did this change touch?

Answers are not in the register. A paragraph written against last quarter's input survives in a slide, a supplier email, a note somebody kept.

Nothing in the quality system knows those exist. The change register reaches every controlled document it is responsible for, and it cannot reach the derived paragraph that left the building weeks ago.

Nobody recalls a paragraph

You can withdraw a controlled document because you know where the copies went. The answer went to one person, in one conversation, and cannot be found again.

The new rules log the tool, not the claim

The European AI legislation requires high risk systems to allow the automatic recording of events over the lifetime of the system. That is a log of the tool, which is useful, and it is not a history of the claim.

The American standards body defines provenance as the chronology of the origin, development, ownership, location and changes of a system and its associated data. A design history file is precisely that for a design. An answer has no equivalent.

Grounded is not recorded

The first preregistered study of retrieval based legal research tools found two leading products returned false information between seventeen and thirty three percent of the time, despite vendor claims of eliminating it. Reading from approved material is not the same as recording what was read.

Give the answer the furniture the file already has

Cognatum is a governed knowledge base. It stores the approved answer itself as a record with a source, a version, a named approver and permissions, so the paragraph an engineer receives carries what the file entry beside it carries.

When a source document changes, Cognatum notifies the people whose answers depended on it rather than quietly rewriting them. Where two approved sources disagree, it surfaces the disagreement and routes it to a person. A human approves. The software decides nothing.

The method already exists here

The gap is not that assistants are unreliable. It is that the organization already knows how to make a decision durable, and has never applied that method to the decisions its software hands out.

Your quality system already governs the file. See what it takes to govern the answers that come out of it.

Common questions

Questions this raises.

What is a design history file?

It is a compilation of documentation describing the design history of a finished medical device. The requirement came in with the Safe Medical Devices Act of 1990. It holds the design plan, inputs, outputs, reviews, verification, validation, transfer and changes, so that how a design was decided can be examined later.

Did the QMSR get rid of the design history file?

It renamed it. Since February 2026 the American device regulation incorporates the international quality management standard by reference, the design controls provisions moved within the rule, and the design history file is now the design and development file. The obligation to document design history did not change.

What has to be recorded in a design review?

The date of the review, the people who took part, the design version or revision that was reviewed, and the results. Verification and validation records carry the same kind of detail, including the method used and the individuals who performed the work.

Why does an AI answer about a design fall outside design controls?

Because it is not an entry in anything. It names no revision, no reviewer and no date against a source, and it is not traceable from design input to design output. When the input changes, the change register cannot reach it, so it is never corrected or withdrawn.

Does logging our AI system cover this?

Only partly. Event logging records what the system did and when. It does not record which approved source an answer rested on, which version that source was in at the time, or who stood behind the answer. Those are properties of the answer, not of the tool.

Knowledge governed. Intelligence everywhere.

See it on your own content, in your own environment.