Search the verified library by discipline, facility type or document number. Every record carries its applicability, revision state and engineering connections — before anyone opens the PDF.
Typed connections between documents are captured once and verified by an engineer. The library then answers the question that costs the most time on a live project — what else has to be reviewed.
Package selection runs on a written, versioned rulebook — not on a model's judgement. The same inputs always return the same approved set, and every document in it can be traced to the clause that put it there.
The rulebook is a controlled document in its own right. Changing a rule is a reviewed change with a version — which is what makes the output defensible in a design review.
The assistant answers only from the verified library, and puts the source clause directly beneath the answer. If it cannot find a grounded source, it says so rather than composing one.
Ask a question, then open a citation to read the exact clause the answer came from.
Nothing enters the library on a model's say-so. Each proposed card and connection arrives with its evidence, a confidence score and a named reviewer — and is accepted, corrected or rejected by a person.
Extraction at scale is cheap; being wrong in an engineering library is not. A card that misstates applicability can send the wrong specification into a live design.
So the pilot is measured on verified records, not extracted ones — and every published record keeps the name of the engineer who accepted it, with the original evidence attached.
Coverage, verification throughput, citation rate and active use — so the pilot is judged on whether engineers actually rely on it, not on whether software was installed.
Engineers report that locating and confirming the current revision of a standard document is a recurring daily cost. The library's purpose is to remove that search, not to replace engineering judgement.
—