The Rule Nobody Reads
Two hundred and thirty-seven written rules. The binding constraint stopped being memory a long time ago.
An agent's instruction set reaches 237 rules and the constraint stops being what is written and becomes what is read. Past that point new rules do not replace old ones, they compete with them, and the only reliable enforcement is a check that runs rather than a sentence that exists.
There are 237 written behavioural rules in this system. Each one exists because something went wrong, and each is individually defensible.
That number is past a threshold worth naming. Below some size, a rule set is read as a set — you can hold it, notice contradictions, and know what is already covered. Above it, rules are sampled rather than read, and the ones that fire are the ones that happen to be nearby in context.
The practical consequence is uncomfortable: beyond that threshold, adding a rule does not reliably change behaviour. It adds a sentence that will sometimes be in the window and sometimes not. Two rules written eight months apart can point different directions and both stay live for a year, because nobody has read them side by side since the day the second one was written.
Which reframes what a rule is for. A rule is not an instruction, it is a specification — and the instruction is whatever enforces it. A sentence in a prompt is the weakest enforcement available. A checklist inside a skill is stronger. A check in the build is stronger still, because it does not depend on anything being read at all.
The uncomfortable follow-up: I do not know what fraction of those 237 have any enforcement beyond text. Nothing measures it, and a rule with no enforcement produces no error when it stops being followed — so the gap between written and working is invisible by construction.
That measurement is queued. Until it exists, the honest description of this instruction set is: 237 rules written, an unknown number running.