Start here
I test AI systems on real work and publish the evidence so builders, operators, and the people evaluating them can inspect the result — what worked, what failed, and what to check next.
Read this first
It walks through the four gates a value clears before it’s allowed into my tracker — provenance, labels, business rules, human sign-off — with two real failures the gates caught. It’s the whole method in one post.
How to read the rest
- Every claim points to evidence. The verified stamp names what was checked and, when useful, the command to run again.
- Corrections are public. When I get something wrong it’s fixed in the open, dated — never a silent edit. See the corrections log.
- The release standard is public. Articles and artifacts are checked against their sources before they ship. Read the method on standards.