Brief / Trust
AI can read a drawing. It cannot approve one.
AI can speed engineering review. Approval stays a scoped, recorded act owned by a named person.
Reading is not authority
Reading a drawing is assistance; approving one is authority. A model may find a title block, compare dimensions, extract a tolerance, or notice that two sheets disagree, but it does not release a revision, establish where it applies, or answer for the result.
Approval is a controlled event: a named person, a defined scope, a specific revision, evidence considered, a decision, and an effective date. The operating system should make that event easier to execute and harder to fake.
Where AI has to earn the right
Veldarium separates assistive intelligence from the control plane. Classification, extraction, comparison, and anomaly detection can become model-assisted inside bounded, tested stages. Permission scope, configuration authority, revision lineage, applicability, conflict, and approval state remain explicit.
The current public readiness proof is deterministic and does not invoke an external language model. Drawing read, compare, check, annotate, and verification-package preparation are planned operating capabilities, not current deployment claims.
The final boundary is a person
An AI-native shipbuilder should make it dramatically easier for engineers and production leaders to see the state of the vessel and the evidence behind a decision. It should not quietly replace the person who is authorized to make that decision.
Human authority is not friction to automate away. It is one of the control surfaces the system has to preserve while everything around it gets faster.