classifier constitution

A classifier constitution is the reference document used to train the monitoring system that decides whether an AI request is dangerous. Just as a court bases its rulings on constitutional text, this monitor learns what to pass and what to block by consulting a rulebook that people have written out in advance.

Writing this rulebook is trickier than it sounds. To make a vaccine, scientists have to cultivate the very pathogen that's dangerous, and some blood pressure drugs come from compounds found in snake venom. Beneficial research and dangerous research can look almost identical on the surface, so if the rules are too loose, dangerous requests slip through, and if they're too strict, they end up blocking ordinary questions from doctors or students. That's why companies gather input from internal and external experts, revise this rulebook repeatedly, and retrain the monitor based on it.

When the rulebook changes, the monitor's entire standard of judgment shifts with it. A question that was blocked yesterday might pass today, or the reverse.

Articles use it like this: "They completely rewrote the classifier constitution that underlies the biological safety classifier's judgments over the course of several weeks." Despite the word "constitution," it's often misunderstood as a legal document, when it's actually closer to an internal training guideline used by safety teams.