A practice already writes things down.
This makes the record answer.

Engineers are the strongest fit for this product, and the reason is not flattering: a practice writes down what it assumed and why because it will have to defend it later. That record is dated, signed and mostly digital already. It is exactly the corpus this engine is built to quote from, and nothing else.

The owner of a warehouse you designed in 2017 wants a mezzanine. The question on the phone is simple: what did you assume for the floor, and can it take more? The answer is in the practice's own record, and the people who wrote it have moved on. The design basis memorandum that stated the loads and the soil. The geotechnical report it leaned on. The calculation package, sealed and dated. The specification. The drawing set, issued as PDF. The transmittal that says which revision went out. The review comment register where the checker argued about the snow drift. The field review report from the day the slab was poured.

The questions you would ask a colleague across the office are exact, because the profession makes them exact. What bearing pressure did we assume on that job? Which edition of the steel code was the 2019 package designed to? Did anyone ever check the culvert at the 1:100 event? What did the reviewer say about the drift load, and did we change it? Each has an answer in a document with a date and a signature on it. The problem is that the document is one of forty thousand, and the file name does not say what is inside.

Strategic Agent reads the practice's documents and answers with a line out of one of them — at most three verbatim lines, at most two citations, each with the document id and the line span. Nothing is generated. You open the file and find the same words at the same lines.

An engineer will notice what that last row is: a bounded claim with its limit stated. The engine can only quote what was converted, the manifest lists what was not converted and why, and a question whose support is not in the corpus gets the refusal instead of a plausible sentence. That is the same discipline as writing "assumed" next to a number.

Three properties of an engineering record make it the right corpus for an extractive engine, and none of them has to be built. The record is digital: calculations in spreadsheets, reports in Word, drawings issued as PDF — the formats it reads. The record is explicit: assumptions are stated because they must be defensible, so the line the answer sits on says "assumed" or "per" or "in accordance with", which is what retrieval finds. And the record is dated and versioned: a transmittal names the revision, so the citation can name it too.

Provenance comes with every answer: the source path, how the text was extracted, the extraction status, and, for a scanned page, the OCR confidence recorded for that page. Originals are never modified; the readable copy is a separate file with its source recorded on it. One thing worth being exact about: provenance is not authorship. The citation tells you which file the line came from and where in it. It does not tell you who wrote the line or whether they were right. That remains your judgement. And the three-line cap is a pointer, not a substitute: it gives you the line and the document, and you read the calculation.

It does not check a calculation, and it does not certify one. It does not do arithmetic by model: where a sum is needed, ordinary code does it and shows the working, and the model only retrieves and quotes. It does not read CAD; a drawing issued as PDF is read as text, and the geometry is not. It does not read old .doc and .xls files yet — there is no reader for them, and 787 of them sit in the founder's own records waiting for one. It runs on the practice's own machine and refuses to start a component that would send your documents to somebody else. There is no language model inside the engine at all: an on-device embedder, and 5,639 lines of Python across 23 modules that you can read.