My research agent cited example.com

This morning my AI research agent confidently cited a study from “example.com.”

I’m building Storyline, a writing app for people whose words are sensitive. Novelists under NDA. Lawyers. Therapists. Journalists protecting a source. The kind of writer for whom “your draft is safe on our servers, trust us” is a dealbreaker.

So Storyline does its AI work on your own device. And when the AI genuinely needs the web, say to research a topic or check a fact, every outbound request goes through one observed door. It screens the request against your manuscript and won’t let a protected name ride along. Then it records, verbatim, exactly what left your machine, on a screen literally titled “What left this device.”

Most apps ask you to trust a privacy promise. Storyline lets you read the receipts.

Getting a small on-device model to actually do this was humbling. For a full day it would pretend to research. It invented plausible citations, fabricated sources, fetched that infamous example.com. It was role-playing the work instead of doing it.

The real culprit was almost funny. The agent’s instructions told it to “summarize the supplied sources.” So it summarized. It never went looking. It didn’t even have a search tool switched on.

Two changes fixed it: tell it to research, and give it the means. The same little model, running locally on a laptop, ran a real web search, pulled a genuine systematic review from Nature and a clinical paper from PubMed, and wrote grounded notes with real citations.

And the ledger showed the single query that left the device: the search itself. Nothing more. No manuscript. No private names.

That’s the whole bet: AI that helps you write without asking you to hand over custody of your words. You shouldn’t have to trust me. You should be able to see.

Still building. More soon.