Standards
Nobody runs a first article on an AI answer
A first article inspection verifies one part against every requirement before a production run is released. The assumption underneath it is that nothing changes. An AI answer breaks that assumption every day, and no control catches it.
Cognatum Team · Nov 18, 2026 · 5 min read
A new setup runs its first part. Before anything else is made, someone measures that part against every line on the drawing, records each result, and signs it. Only then does the run open.
Cognatum governs the entry
source · version · approver · permissions
It is a plain idea, and an old one. Prove the first one before you make the rest.
Nothing like it exists for the answers your people get from an AI assistant. The first answer after a source changes goes straight to whoever asked for it.
No one measured it. No one signed it. There was no first article.
What a first article proves
A first article inspection is a documented check that a new or modified production process turns out parts matching the specification in the engineering drawing. The supplier performs it, and the buyer reviews the report.
Every characteristic, numbered
The aerospace format breaks the record into three forms, and the third one is the heart of it. Characteristic accountability. Every dimension, material property, special process and drawing note gets a balloon number of its own.
Against each number sits the requirement, the measured result, a pass or fail, the inspection method and the identity of the gauge used. Hundreds of lines, with nothing left unaccounted for.
The assumption underneath
One part is enough because of an assumption that usually goes unwritten. Every part after it will come off exactly the same process. Inspect the first, and you have said something about all of them.
Any change to the design or to the process breaks that assumption. The guidance is explicit about it. When the assumption breaks, you run a new first article.
The trigger list is the interesting part
What makes the control work is not the measuring. It is the list of events that force a fresh check. A new part. A design change affecting fit, form or function. A change of process, tooling, material, supplier or location.
There is one more trigger that has nothing to do with change at all. A lapse in production of around two years reopens the file, even when the drawing and the process are untouched.
Time alone is a reason to look again
That is a quality system conceding something useful. Confidence decays on its own. If enough time passes with nobody watching, the right move is to verify again rather than to assume.
Partial checks, only where it moved
The standard also allows a partial inspection, often called a delta. When a change is limited, only the affected characteristics get fresh data. Everything untouched stands on the earlier record.
That is a precise instrument. Something changed, these are the characteristics it reached, re-verify those and leave the rest alone.
Why an answer is never a first article
There is no stable process to point at
A production line can be held still. A knowledge base cannot. Policies get revised, contracts get amended, a procedure picks up a new revision, somebody saves a copy and edits it.
Retrieval shifts as documents are added. Model versions change underneath. Every answer is effectively the first one off a new setup, so inspecting any single answer says nothing about the next.
Nothing is numbered
An answer arrives as a paragraph. There is no characteristic list, no requirement column, no recorded result, no method and no instrument identity. Nothing in it can be checked line by line, because nothing in it is a line.
The part has already shipped
A first article happens before the run is released. That timing is the whole control. An answer is produced and delivered in the same moment, so there is no before in which to inspect it.
The person who asked has already acted on it. Any check that exists is one the reader performs, silently, using judgement they may not have.
What the rules reach, and what they miss
Quality management requirements govern documents, processes and records, and they do that well. The gap is not negligence. It is scope. These systems were written for artifacts that sit still long enough to be inspected.
Accuracy is declared for the system
European AI rules require a high risk system to declare its accuracy level and the metrics behind it in the instructions supplied with it. That is a statement about the tool, published once.
It is not a statement about the answer in front of a particular person on a particular morning. A declared accuracy rate is a population figure, and a reader needs to know about one case.
Grounding is not an inspection
A preregistered study of commercial legal research assistants built on retrieval found that tools marketed as eliminating fabrication still produced it between seventeen and thirty three percent of the time.
Pointing a model at approved material improves the odds. It does not verify the output. Those are two different jobs, and only one of them has been done.
What an answer would need
Provenance, in the security literature, is the chronology of the origin, development, ownership, location and changes to something and its associated data. That is a characteristic list, written for knowledge instead of for a casting.
A record for each answer
An answer can carry the furniture a part already has. A named person who approved the underlying entry. The date they did it. The source it came from, at the version that was live when the answer was given.
A trigger when the source moves
The delta inspection has a direct equivalent. When a source document changes, the useful question is narrow. Which approved entries drew on it, and which answers depended on those entries.
Cognatum watches the sources it is pointed at and raises a notification when one of them changes. It does not quietly rewrite anything, and it does not claim to be current on its own. It tells a person what moved, and what that movement touched.
A person still signs
The Cognatum Knowledge Loop runs eight steps across four phases. AI does the finding, the structuring, the cleaning and the surfacing. Approve stays with a named human, who can rework an entry before putting their name to it.
Where two documents contradict each other, both are presented with the disagreement marked, rather than one of them being silently chosen.
Prove the first one
A first article inspection is a quality system saying that the first one off a changed process must be proved before anyone trusts the rest. That sentence has never been written for knowledge.
Your company's knowledge isn't missing. It's unusable. Cognatum changes that.
Sources
- A Guide to First Article Inspection (1factory.com)
- First article inspection (wikipedia.org)
- AS9102 Aerospace First Article Inspection Requirement (sae.org)
- First Article Inspection, including partial and delta FAI (industrial-ia.com)
- ISO 9001:2026 Quality management systems, Requirements (iso.org)
- EU AI Act Article 15, Accuracy, Robustness and Cybersecurity (artificialintelligenceact.eu)
- Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (arxiv.org)
- Provenance, Computer Security Resource Center glossary (nist.gov)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
What triggers a new first article inspection?
A new part, a design change affecting fit, form or function, a change of process, tooling, material, supplier or manufacturing location, and a lapse in production of around two years. Each one breaks the assumption that the next part comes off the same process as the one already inspected.
Is a first article inspection the same as validating an AI system?
No. System validation asks whether the software behaves as specified. A first article asks whether one output, produced after something changed, matches every requirement on the drawing. Most AI assurance work sits on the system side, which leaves individual answers unexamined.
Why not simply inspect a sample of AI answers?
Sampling rests on a stable process, which is what makes one part representative of the rest. A knowledge base changes continuously, so an answer sampled on Monday tells you little about the same question asked on Friday against a revised source.
What is a partial or delta inspection, and what is the equivalent for knowledge?
A partial inspection re-verifies only the characteristics affected by a change and leaves the rest on the earlier record. The equivalent for knowledge is knowing which approved entries drew on a document that has just been revised, and which answers depended on those entries.
Does Cognatum keep knowledge up to date on its own?
No. Cognatum notifies when a source it is pointed at changes, and shows what that change affects. The decision to revise and approve stays with a named person, which is the Approve step of the Cognatum Knowledge Loop.