A quality-control laboratory is replacing its laboratory information management system. Three suppliers have responded. The validation lead has a requirements specification of 140 lines, written over a fortnight by six people, in a spreadsheet.
Line 62 reads: "The system shall maintain an audit trail."
Everyone agrees with line 62. Nobody can say, eight months later, why it was a must-have rather than a should, which supplier's design actually satisfied it, whether the test that checked it was the test that was run, or what happened to the deviation raised when one of the tests failed.
Each of those answers existed at some point. They were in four different documents, owned by four different people, connected only by the fact that the same sentence appeared in each.
This is the traceability problem, and it is not a paperwork problem. It is a problem of things that should be joined being kept apart.
Four Things That Should Be One Chain
In any regulated delivery, four kinds of statement matter.
The rule
What the regulation and your policy demandThe requirement
What this system must doThe design
What was chosen to do itThe test
What proved it didFigure 1 — The chain an inspector follows, in either direction.
An inspector reads this chain backwards. They pick a test result and ask what it proves. They pick a requirement and ask why it exists. They pick a clause of a regulation and ask where in your system it is satisfied.
Most organisations can answer each question separately. Very few can answer them as one chain without a week of preparation, because the rule is in a quality manual, the requirement is in a specification, the design is in a supplier's slide deck, and the test is in a validation tool.
cajeX already held the first link. A regulation in the framework catalogue, such as 21 CFR Part 11 or EU GMP Annex 11, is broken into clauses, and each clause is answered by one of your approved directives. What follows is the rest of the chain: Requirements, Design, and Verification, which a workspace owner or admin turns on in Settings → Features.
Link One: A Requirement Knows Where It Came From
A directive is standing policy. It is written once and applies to every project: "Electronic records that support a GxP decision carry a secure, time-stamped audit trail."
A requirement belongs to one solution: "The LIMS records who changed a test result, when, the previous value, and the reason, and that record cannot be edited or deleted by any user role."
The second exists because of the first. In cajeX that is a recorded link, not an assumption. On the requirement's Traceability tab you see what it derives from, what satisfies it, and what verifies it.
21 CFR Part 11 §11.10(e) · EU GMP Annex 11 clause 9
Audit trails for electronic records: Part 11 for actions that create, modify or delete them; Annex 11 for changes and deletions.
GxP Records Carry a Secure Audit Trail
Approved policy. Applies to every system that holds GxP records.
CR-0062 · Must
The LIMS records who changed a result, when, the previous value and the reason.
Derives from: GxP Records Carry a Secure Audit Trail
Figure 2 — Line 62, with its provenance attached.
That provenance does practical work. A requirement that derives from one of your own directives is one a supplier cannot negotiate away. When cajeX exports a vendor pack, the set of requirements a supplier is asked to meet, those requirements are marked non-negotiable, so the supplier knows before pricing the work.
A few details keep the set usable at 140 lines.
Each requirement has a reference such as CR-0062 that is unique across the workspace and never changes, even if the requirement's type does. By the time you might want to renumber it, you have quoted it to a supplier.
Priorities are MoSCoW, and only Must is emphasised. If everything looks urgent, nothing does.
Each requirement carries acceptance criteria, which answer the question "how would you know this is met?". Write them even when nobody is testing yet. They are what tests are later drafted from.
You can draft a starting set with AI. It reads the project's description, its documents, and above all the project's directive scope, so requirements are proposed against what this project actually has to conform to. If the model refers to a directive that is not in scope, cajeX discards that link and tells you how many it dropped. A generated requirement is never shown as policy-derived unless it is.
Link Two: Candidates Are Judged Against the Requirements
Three suppliers means three designs. In cajeX each is a candidate on the project's Design tab, with its own architecture views at levels from business context down to deployment.
The Compare view puts requirements down the side and candidates across the top. Each cell is a verdict: Met, Partial, Not met, or not applicable. A Partial or Not met verdict asks what is missing, because by the time a supplier disputes it, whoever judged it will not remember.
| Supplier A | Supplier B | Build in-house | |
|---|---|---|---|
| CR-0062 Audit trail on result changes | Met | Partial — reason for change is optional | Met |
| CR-0071 Backups restorable and tested | Met | Met | Not assessed |
| NFR-0018 400 concurrent analysts | Partial — tested to 250 | Met | Not assessed |
| Must-haves met | 31 of 38 | 33 of 38 | 12 of 38 |
| Must-haves not yet judged | 2 | 0 | 24 |
| Policy-derived requirements unmet | 0 | 1 | 0 |
Figure 3 — The comparison, with the two rows that are deliberately kept out of the score.
Look at the last two rows. They are there because a single score hides the two things that matter most.
A candidate can look strong on what has been examined while most of the set has never been considered. "Build in-house" scores 12 of 38 here, but 24 have not been judged. That is not a weak candidate; it is an unexamined one, and the table says so.
And missing a requirement that came from your own directive is a different kind of failure from missing a preference. Supplier B is ahead on the count and has one policy-derived requirement unmet. That line never disappears into a percentage.
You can ask AI to propose verdicts for a candidate. Every proposal must give a reason or it is thrown away, proposals stay marked as AI until a person confirms them, and your own judgements are never overwritten. Where the supplier's material is silent, nothing is recorded: marking silence as Not met would hold a supplier responsible for a gap in your own evidence.
Selecting is a decision, and it is recorded as one
Select this design turns a candidate into the decision. Three things follow, and the confirmation states all three before anything happens.
The other candidates are rejected and kept as the record of how the choice was made. A draft architecture decision record is written on the application, using the reason you typed, and left for you to complete and accept. And every unmet must-have opens a finding: a nonconformity where the requirement came from a directive, a risk otherwise.
That last step is the one that prevents the usual failure. Choosing Supplier B with the audit-trail gap is a legitimate decision. Forgetting that you did is not.
Link Three: A Test Names What It Proves
A requirement says what must be true. A test says how it will be proven. In cajeX a test names the requirement it verifies, so a passing result is evidence about something specific.
Each test carries an objective, a method (executed test, demonstration, inspection, or analysis), numbered steps, and an expected result that is observable. "Works correctly" is not a result. "The audit trail shows the previous value 4.2, the new value 4.7, the analyst's user ID and the reason entered" is.
Tests can be drafted with AI, one per requirement, from the acceptance criteria. Protocols can be drafted against the selected design in the three stages a regulated handover expects.
IQ
Installation
OQ
Operational
PQ
Performance
Is it installed and configured as designed?
Does it operate as specified under normal conditions?
Does it perform under production-like conditions?
Figure 4 — The three verification stages. A workspace chooses the vocabulary: IQ, OQ and PQ for GxP, or the plain-English, enterprise IT or plant equivalents.
The stages always ask the same three questions. What they are called is the workspace's choice under Admin → Terminology, because a supplier in pharma expects IQ, OQ and PQ, while a plant engineer expects site acceptance, commissioning and a performance run. Changing the vocabulary changes only the words.
Everything AI drafts lands as a draft, and a draft is a test nobody has read. Only tests marked ready are included in the protocol pack, the document a supplier receives. If nothing is ready, the pack says how many drafts are waiting.
What "verified" means
This is where most traceability matrices mislead.
Verified
Every linked test has actually passed
Failing
At least one linked test's latest run failed
Not yet proven
Tests are linked, but not all have passed
No test
Nothing is linked yet
Figure 5 — Four verification states. "Not yet proven" is kept apart from "Verified" on purpose.
A requirement with three tests, where one passed and two were never run, is not verified. Calling it so is how a gap reaches production wearing a green tick. Blocked and skipped runs prove nothing and count the same as never run.
The state is worked out from the latest result of every linked test each time you look. It is never stored, because a stored copy would be wrong the moment someone recorded a result.
Results are events, not edits. Re-running a test adds a result and keeps the old one, so the history of what failed and when survives. A failed result offers to raise a finding, pre-filled with the test, what it verifies, what happened and what was expected, and the finding's ID is written back to the result. The trail runs both ways.
Must-haves with no test at all are listed by reference. A failure is something you know about. An untested must-have is something nobody has thought about yet.
After Go-Live: The Chain Stays With the System
A project ends. The system is what you still run on Monday.
When the project completes, its requirements move to the applications they were allocated to. They move, they are not copied: same reference, same status, same link to the directive. They become the application's standing requirements, what must stay true whichever project touches it next. The selected design and its views are promoted to the application in the same way.
The next project to touch the LIMS sees those standing requirements at the top of its own Requirements tab, shown but not editable. CR-0062 is still there, still derived from the audit-trail directive, still linked to the tests that proved it.
Kept apart
- Rule in the quality manual
- Requirement in a spreadsheet
- Design in a supplier deck
- Test in a validation tool
- Joined by a sentence that matches
One chain
- Clause answered by a directive
- Requirement derived from it
- Design judged against it
- Test that names it
- Joined by recorded links
Figure 6 — The same four statements, kept apart and joined.
What This Is Not
It would be easy to overstate this, so here is the boundary.
The protocol pack is a grouped test pack. It is not a controlled validation protocol. It has no approval signatures, no locked versions and no deviation log. That apparatus belongs to a validation execution system, and half-building it would look like compliance without being it. cajeX carries the traceability that feeds yours.
Nor does any of this make a system validated. Validation is your quality unit's judgement, made under your procedures. What cajeX provides is the thing that judgement depends on and that is usually missing: for every requirement, where it came from, what was chosen to meet it, and what was run to prove it, in one place and in order.
For the regulations themselves, the catalogue includes 21 CFR Part 11, GAMP 5 Second Edition, EU GMP Annex 11 and Annex 15 among the frameworks a workspace can enable. See regulated industries for how these fit together, and one review as evidence for every regulator for what happens when several of them apply to the same decision.
Line 62 is still one sentence. It is no longer four copies of one sentence in four places.
Sources
- 21 CFR Part 11 — Electronic Records; Electronic Signatures, US Food and Drug Administration, §11.10(e)
- EudraLex Volume 4 — Good Manufacturing Practice guidelines, European Commission, Annex 11 (Computerised Systems, the January 2011 revision, clause 9) and Annex 15 (Qualification and Validation)
- GAMP 5 Guide: A Risk-Based Approach to Compliant GxP Computerized Systems, Second Edition, ISPE, July 2022