I wrote a set of first principles for the system I run at home. Principle 9 is simple: store facts, derive views. Keep one canonical source of truth; calculate summaries and interpretations from it rather than freezing them into new copies.
Then I measured what my agents actually load at the start of every session.
The principles file was 6,444 bytes. The behavioral rules layered above it were 24,172 bytes — much of that spent restating the principles instead of pointing back to them.
Principle 9 had become a 24 KB example of breaking Principle 9.
The cost was bigger than wasted context. Every restatement created another place for the doctrine to drift, contradict itself, or survive after the original principle changed. More "enforcement" was making the system less coherent.
The fix wasn't another rule. It was subtraction.
The principles remain the source of truth. Behavioral rules now point to them instead of copying them. Auto-memory holds only what can't be derived from a principle: exact literals, hard bans, explicit exceptions, host-specific gotchas.
One fact, thin references, behavior derived at read time.
It sounds like a small cleanup, but it's the kind of defect a design review rarely catches. You find it by running a system long enough to watch its implementation drift away from its doctrine — then measuring the gap.
Still tightening it.