Law Outranks Intelligence. Including Ours.

I wrote that sentence in a Hacker News comment five months ago and it got zero engagement, which is fair, because I buried it at the bottom of a feature list. It deserves better, because it's the entire design philosophy of CORE in five words.

Here's the problem it answers.

Every AI coding setup eventually faces the same question: when the AI is smarter than the check, who wins?

Your linter flags a pattern. The LLM explains, persuasively, why this case is an exception. Your architecture rule says no cross-layer imports. The agent has a genuinely clever reason why just this once is fine. And if you're honest, sometimes it's right. The model reasons about the codebase better than the rule does.

So the temptation is to let intelligence adjudicate. Give the smartest component the final say. Make the rules advisory and the reasoning sovereign.

That's the trap. A system where intelligence outranks law has no law. It has suggestions, and a very articulate entity that decides when they apply. Every "justified exception" is indistinguishable — in the audit trail, six months later — from a violation with good marketing.

CORE inverts this. The constitution lives in a directory called .intent/ — rules as data: YAML, checked by deterministic engines, enforced with a hard halt. The autonomous daemon that governs the repo cannot write to that directory. Not "is instructed not to." Cannot — the write path is blocked at analysis time, at runtime, and at commit time, by three independent stages.

But here's the part people miss, and the reason the sentence ends the way it does.

I can't bypass it either.

I'm the governor. I hold root on the machine, admin on the repo, and authorship of every rule. And when I want to change a rule, I go through the same governed path: a draft, a review, an audit trail entry with my name on it. The constitution doesn't know or care that I'm the smart one in the room today. When my own commits violated docs-integrity rules, the audit flagged me. I spent a Saturday fixing my own repo because the system I built made me.

That asymmetry - law above every intelligence in the system, artificial or human — is what makes the audit trail mean something. If any actor can adjudicate exceptions on the fly, the trail records outcomes, not authority. You know what changed but not what permitted it. In a regulated environment, that second question is the only one the auditor asks.

The practical consequence is almost boring: the smartest component in CORE (an LLM-as-judge engine that reasons about intent) runs in reporting mode until a governed decision promotes it. Intelligence proposes. Law disposes. The most capable reasoner in the system has exactly as much authority as the constitution grants it, and not one merge more.

People building agent systems keep asking how to make the AI more aligned, more careful, more trustworthy. Wrong layer. Make the system one where trustworthiness isn't a property of the actor at all. Then it doesn't matter how smart the actor gets.

Law outranks intelligence. Including ours. Especially ours.

CORE is open source: https://github.com/DariuszNewecki/CORE