It looks things up. And it says when it could not.
A search that failed is not a search that found nothing. Those are different facts and they never arrive as the same sentence.
Research
A search that failed is not a search that found nothing. Those are different facts and they never arrive as the same sentence.
What holds
The refusal that matters
It cannot give you a number it cannot source.
A figure that is not present in what it read is dropped, and a figure it derived is allowed only when one named piece of arithmetic over the evidence produces exactly it - with the working available. The whole answer goes rather than a bad number being quietly stripped from a sentence that would no longer mean what it says. (ground.ts, derive.ts)
A failed search says so
No key, no reach, a challenge page it could not pass - each is reported as itself. "I found nothing" is only ever said when it looked and there was nothing. (search.ts)
Four sources, four failure modes
A plain fetch, a real browser, an encyclopaedia API, and a paid index. They fail for different reasons, which is the only kind of redundancy worth having.
Nothing it reads gains standing
Something read on a page is marked as read, and passing through the agent cannot promote it to something you said. (provenance.ts)
It is not paid to prefer anything
No result is ranked by anyone who paid. There is nowhere in the code for that to enter. (loyalty.ts)
How it goes
From the question to the answer, or to the reason there is not one.
You ask in passing
No tool to pick, no mode to switch into.
It looks, on your day’s budget
Bounded, and the trip is recorded like every other departure.
It reports what it found
With where it came from, and what it could not confirm.
Or it tells you it failed
Plainly, and with the reason, so you know whether to try again.
Until you ask
It does not go looking on its own. So this stays empty until you ask.
No background crawling, no speculative lookups, nothing running up a bill while you sleep. A search happens because you asked a question that memory could not answer, and for no other reason.
No lookups yet
When a question needs something it does not already know, it goes and looks - and every source it used is listed here, including the ones that turned out to be no good.
What it holds to
Four things it will not do to sound certain. Fluency is not knowledge.
It never fills a gap with fluency
A model asked for a figure will produce one. The grounding check is what stops that reaching you.
It never hides which source answered
You can see whether it read a page or an API, and whether the page challenged it.
It never sends you with it
Your name, your company, your number stay here unless the question needed them - and if one travelled, it is in the ledger.
It never rounds unknown to none
"I could not verify this" is an answer it is willing to give.
The rest of the workbench.
Decisions arrive with the work done, not the questions still open.
Context that stays warm after the meeting, the email, and the tab close.
Meet the chief of staff your company never had.