The tool call

The demo above runs the same question down both roads. Ask in text and the digits come out of the odds — answer-shaped, maybe right, maybe not, and identical-looking either way. Flip to the tool path and the model writes a request instead; the app runs it as real code; the true product comes back as text into the window, and the reply continues from it.

That is a tool call, and it is the general mechanism behind every "the AI checked something" moment: web search, code, file lookup, calendars, databases. The model never touches any of them. It writes text that asks; the app executes; the window receives what came back. One door, three steps: ask → execute → paste back.

Two consequences worth owning:

• Performed vs narrated. "I ran the numbers" is a sentence, and sentences are cheap for a pattern machine. The app's visible tool step is the receipt that something actually ran. No receipt means you can't credit the check — treat it as completed, not computed.
• Much of what feels like a smarter AI is a better-plumbed app. Same mechanism, more doors — search, code, files, calendars wired in. When a product seems to "know" something no model could, look for the pipe.

Say it back in one line: why do a computed answer and an invented one look identical in the reply — and what, outside the reply, tells them apart?