Start here

A plain guide to the site, for readers with no background in AI.

In short

This site holds the work of Zain Dana Harper, published under the working name Zentropy Labs. The work keeps returning to one question: when an AI system, or the company behind it, makes a claim, how can someone else check it? Pick the door below that fits what you came for.

Four ways in

If you came to read

Who Knew First traces nine AI agent incidents from 2026 and asks how long the public waited to hear about each one. It opens with a short op-ed and keeps the full record underneath.

Read Who Knew First

If you follow AI policy

The Frontier Safety briefing is a dated record of safety news. Each edition ties every claim to its source and says what that source cannot show.

Read the current briefing

If you want to try the tools

Flywheel runs a task with any AI model and hands the result to a checker that anyone can rerun offline. Install it with one command.

See Flywheel

If you want to hire or work together

The work page lists resumes, the full CV and ways to get in touch.

Go to the work page

What re-derivable means, in one example

Say a company reports that its model passed 120 of 150 coding tasks. A re-derivable report comes with the tasks, the model version, the test that graded each answer and a record of every result. You run the same test on your own computer. If you also get 120, the verdict is match. If you get a different number, the verdict is drift, and the record shows the exact task where the results part ways. If the report leaves out something you need, such as the tasks themselves, the verdict is unverifiable, and the missing piece is named.

One claim, one rerun, three possible verdicts A company's claim of 120 of 150 tasks passes to your rerun of the same test. The rerun leads to one of three verdicts: match if you also get 120, drift if you get a different number, unverifiable if the record is missing a piece you need. The claim 120 of 150 passed Your rerun same tasks, same test Match: you get 120 Drift a different number Unverifiable a piece is missing
One claim, one rerun, three possible verdicts. How to read this: follow the arrows from the claim to your rerun, then to the one verdict your result lands on. The filled box, the plain box and the dashed box carry the three outcomes, so the diagram reads without color.

How to read the charts and diagrams

Each chart opens with one sentence that says what it shows, then a short line on how to read it. Labels sit on the chart itself, in plain words. Under each chart, a section called How we know holds the sources, sample sizes, dates and limits. Open it when you want to check the work.

How AI assistance is labeled

Much of this site was written with AI assistance, and the pages say so where it applies. The prose is checked with Articulate, a writing checker built here, before it goes up. A clean check means the writing avoids known patterns of filler, hedging and stock phrasing. A clean check says nothing about whether a claim is true. The sources and the How we know sections carry that weight.

What this site does not claim

The tools are built to serve people who evaluate AI systems. No pilot, contract or engagement with any AI lab is in place. Where a page reports a result, it also says what the result cannot show.

Words you will meet

A short glossary explains the terms this site cannot avoid, such as receipt, verdict and witness.

Open the glossary