Menu
Before any of it works

Ask twice, get the same answer.

Most of what goes wrong with AI on a jobsite happens before the answer: the documents weren't really read, or they get re-read differently every time you ask. Both are fixed at the front door.

Read once

At the door, not at the question

Your specs get broken into their own paragraphs the moment they arrive — 2.1.B stays 2.1.B. Every answer later cites a paragraph you can find in a paper copy, not a floating chunk of a page.

Same answer

Tuesday and Thursday agree

Ask the same question twice and you get the same answer back, marked as the answer of record, with the date it was first given. Told 5/8″ on Tuesday and 1/2″ on Thursday is how a wall gets torn out.

But not stale

A change kills the record

New revision, an ASI, an RFI landing on that section — the recorded answer dies and it's worked again from scratch, and it tells you which of those happened. Consistency that outlives the documents is worse than none.

Scans

A copy of a copy still reads

Old sheets, marked-up prints, anything that went through a copier has no text in it at all. Those get read off the image — and the words stay pinned to where they sit on the page, so you can still see the line.

Bad scans

Read twice, then reconciled

Schedules, two-column pages, small type on a reduced print — the pages a character reader gets wrong get read a second time a different way, and the two readings are checked against each other line by line.

The hard rule

Nothing gets added that isn't there

A line that can't be traced back to the scan is thrown out and counted. An invented schedule row reads perfectly, which is exactly why it can never be allowed to become a requirement.

RECEIVED what arrived, and when READ text, or off the image BROKEN OUT into numbered paragraphs INDEXED standards, ratings, sizes SHOWN TO YOU before you rely on it

Open any document and you can see exactly what came out of it — how it was read, how many paragraphs, how many carry a “shall,” and every line a second reader tried to add that didn't match the scan. A half-read spec book should never look the same as a fully read one.

Every answer

Trust, but verify. So we show you the page.

Russ doesn't ask you to take his word for it. Every answer carries the document behind it — the actual page, with the exact line highlighted.

See it

The actual page

Click through and you get the page as it was uploaded, with the referenced lines boxed. Not a summary. Not a quote he typed. The page.

Drawings

A snippet of the sheet

A 30×42 sheet on a phone is a grey rectangle. So a drawing comes back cropped to the area in question and enlarged, with the whole sheet one tap away.

Scaling

He tells you he scaled it

Scaled off a drawing, interpolated, assumed — it says so, shows how he got there, and drafts the RFI. Scaling is fine. Scaling quietly isn't.

Sign-off

A person, not a machine

Anything he worked out stays unvalidated until someone checks it and says what they checked. Your name and the time go on the record.

Missing

“Not in your documents” gets proved

He names the section that should have covered it and whether you've uploaded it. If a document is on the shelf but couldn't be read, that isn't not there — it's we can't tell, and he says that instead.

Outside

Off-contract answers get labelled

Ask him a general question and he'll answer — marked as coming from outside your contract, and whether he looked it up or already knew it. The spec still governs.

CITED backed by your documents DERIVED he worked it out · needs a human OUTSIDE not your contract NOT ON FILE and here's where it should be

Every answer is recorded — the question, what backed it, who signed it off. If a reviewer ever asks how you knew, the answer is a document, not a memory.