Prompts advise, types describe, gates enforce
A rule written in a prompt holds until the model has a good enough reason not to follow it. The mechanism hierarchy, and why every gate must say what it fails to catch.
Essays on data-centric systems, designing software as a graph, and keeping autonomous agents honest. Published here first, carried to your inbox by the Esoteric AI newsletter.
Newest first. Essays that belong to a series are also collected, in reading order, under series.
A rule written in a prompt holds until the model has a good enough reason not to follow it. The mechanism hierarchy, and why every gate must say what it fails to catch.
Agents multiplied what a codebase can become without multiplying anyone's ability to notice. Ashby's law, Beer's Viable System Model, and drift as a variety imbalance.
Data-centric backends, SemanticSeams and regulator look like three businesses. They are one idea, applied to data, to software design and to autonomous agents.