A German Apple magazine reviewed one of my side projects this month, and the part that stuck with me was a single throwaway sentence in the middle of the test.
Mac & i, issue 4/2026, page 114, reviewed my Safari extension Trackless Links. The verdict was kind. But the detail I keep thinking about is that the reviewer noticed the credibility indicators the app shows for news sites, and noted that the extension checks addresses locally against an open list of roughly 2,673 entries.
KI fรผr Content Creation
Entdecken Sie das neue KI-Buch zu ChatGPT, DALLยทE, MidJourney & Co.
Amazon Mehr erfahrenThat list is not a licensed feed and not an API call to a vendor. It is a research dataset. How it got from an academic repository onto someone’s iPhone is a better story than any review.
The problem the feature is trying to solve
My research deals with online disinformation and how people judge whether a source can be trusted. One finding keeps showing up across that literature: the decisive moment happens before reading, not after.
Once an article has been read, the impression is set. Corrections reach a fraction of the people who saw the original claim. The countermeasure researchers keep pointing to is pre-bunking: context arrives before the content lands, rather than as a correction weeks later.
That is a tidy argument to make in a paper. Acting on it requires something distinctly less glamorous: a reliable, auditable list of which domains have a track record of trouble. That is exactly where things stall.
Why CRED-1 exists
Sources for this kind of judgment do exist. Most of them carry at least one of three problems. They are proprietary, so nobody can check the reasoning. They are API-bound, which means every lookup reveals which page someone is reading. Or they are an undocumented list with no account of how it was assembled.
For research work, all three are disqualifying. Reproducibility is not paperwork, it is the foundation. So I built an open dataset instead: CRED-1.
CRED-1 merges several openly licensed source lists with computed signals and assigns each domain a credibility score between 0 and 1. The inputs include assessments from existing fact-checking projects, domain age, reach measured against the Tranco ranking, documented fact checks of published claims, and security warnings.
What mattered most to me was that every individual score stays inspectable. Each domain carries its component values with it, so you can see why a number came out the way it did. The dataset is CC BY 4.0, has a DOI on Zenodo, and lives in full on GitHub. It currently covers 2,674 domains and grows weekly. The 2,673 the review mentions was the count at testing time.
I presented CRED-1 at ACM WebSci 2026. The accompanying paper, written with Martin Kappes and Marc-Oliver Pahl, is available as a preprint.
From dataset to extension
A dataset on Zenodo is exactly that: a file other researchers can cite. Valuable, but it changes nothing for the person opening a forwarded link on the couch at night.
That gap between “published” and “useful to somebody” has bothered me for a while. Trackless Links already existed as a tool that strips tracking parameters out of URLs. An extension that inspects every address you open anyway turns out to be the natural place to surface a note about the source.
The implementation was the part that mattered. The obvious approach would have been a server the app queries on every page load. Technically trivial, conceptually absurd: an extension that removes tracking while shipping a complete browsing history to a server would be its own opposite.
So CRED-1 ships entirely on the device. Matching happens locally, no request leaves the phone, and I have no idea which sites anyone visits. The cost is that the dataset updates with app releases rather than in real time. For a use case where assessments shift over weeks rather than seconds, that is the right trade.
The feature is off by default, deliberately. A note about a source is context, not a gate, and nobody should get an unrequested verdict pasted over their reading. Turning it on is a choice the user makes.
What I take from it
Getting from a research question to an open dataset to a shipping app took longer than I expected, and the unpleasant part was not the science. It was maintenance. A dataset published once and never touched again decays fast.
A trade magazine mentioning the feature in passing is, to me, the actual result. Not because it is coverage, but because it means the work left the academic context and now does something for people who have never heard of the underlying research. That was the goal.
CRED-1 is open and usable via npm or as an MCP server if you want to build something with it. Missing or misclassified domains can be reported directly on GitHub. I would rather have either of those than the review.
The full review is in Mac & i 4/2026 (paywalled, German). Trackless Links on the App Store. A German version of this article is also available.
