Document review normally reads what ARRIVES. This reads what is about to LEAVE. The asymmetry is the whole point: by the time a request has gone out, the cost of the mistake inside it is already fixed.
The instant pass runs on the wording alone. It costs nothing, returns in a second, and is safe to run while somebody is still typing. The full check spends real effort: it goes and retrieves the specification sections the draft leans on and reads them whole, which is the only way to catch a contradiction nobody remembered.
These run once per specification revision rather than once per question, and they are the reason the check can be specific. A conflict found at tender is a priced qualification. The same conflict found at procurement is a delay and an argument. And a clause that already names your manufacturer as acceptable is the strongest sentence your letter can carry — most of them are never used, because nobody knows the clause is there.
Substitutions are never local. Change one fitting's thread and you have quietly obliged new couplings, nozzles, reducers, caps and the fire authority's inlets — none of which appears in the request that started it. Mixed threads in a tower is a hazard discovered at commissioning, which is the worst possible moment to discover anything. The workspace walks the dependency and lists what else the change touches: other components, approved submittals, procured items, drawings, other trades.
A project carries years of clarifications, each with a written response from the party who will answer the next one. That is institutional memory sitting as files, findable only by whoever was there at the time. The workspace turns it into an answer: this has been ruled on, here is the ruling, here is how it was reasoned — and it predicts the likely reply from what this Engineer actually wrote, not from a generic model of consultants.
Everything above stops a mistake leaving. The record is what protects you when the answer is slow.
Two places the same evidence goes further: out to the site, and up to the head office.
A polished demo proves nothing. The honest test of a document engine is a trick question, asked on a real specification. Bring a sanitized project to your trial and put these three traps in front of it:
Ask about a requirement where two clauses of the same specification disagree. A keyword tool returns the sentence that agrees with you. The check is built to read the section whole and surface the one that does not.
Cite a section number from a code edition your contract does not bind. The editions register exists exactly for this: the rejection that looks like a technicality, caught before it costs a full cycle.
Ask a question whose answer a previous Engineer response quietly reversed. The precedent search answers from your own correspondence, with the letter and the date.
If it misses one, tell us exactly what it missed. That report is worth more to us than a compliment, and it is how the engine got this far.