How I work
I run my projects on an operating system for AI agents that I built and published: DIE-OS. It is a way of working, not a product. Nothing in it knows what I am building.
Three files
Roles, gates, laws. Fourteen constrained expert roles, each with an operating prompt, a checklist that gates its output and a memory that fills up with lessons in the project. Fourteen gates every artifact passes before a human is asked to approve it. Conduct laws, append only, each named after the defect that caused it, with a script that checks the ones that can be checked mechanically, because prose decays and scripts do not.
One judge
For anything that claims to measure a feeling, the ground truth is my verdict, given blind. Agents build exams, train models and write reviews of each other's work; none of them decides what a feeling is.
Pre-registration
Before money is spent on a training run, the bars are frozen, the critical counts are written down, the number of times the exam may be read is fixed, and an adversarial reviewer files its own prediction. A result is read against what was written before it existed.
Refusal built in
The system is allowed to refuse me. A deck builder will not write a deck that breaks the composition law. A training job will not start without the reviewer's acknowledgement in its provenance file. A reviewer rewrote my own claim about a result this week, and it was right to.