THE COMPLIANCE COSTUME

Most AI governance tooling records what happened, then dresses up as a tool that governs what should. A self-taught systems architect who builds off the cloud told me why that costume falls apart the moment you leave it.

Montgomery Kuykendall is the founder of Kuykendall Industries, a self-taught systems architect who builds AI that runs locally, off the cloud. He came at governance through the side door, and that matters more than it sounds. The credentialed path trains you to trust the form, the dashboard, the sign-off. Come in sideways and you never learn to stop asking who in the room could be lying.

He reviewed the piece before publication. The quotes below are his words, lightly trimmed for length. The framing around them is mine.

It was never governance

I asked him where today’s tooling breaks the moment you leave the public cloud.

“Most of it breaks pretty much immediately, because most of it was never governance tooling. It was cloud observability with a compliance costume. They’re not really governing, they’re observing. It breaks because it assumes platform accounts exist. Logging breaks because vendor logs are not portable. Reproducibility breaks because nobody records the full decision tree and chain. Tool traces break because they treat it like a webhook and not a path or a flow. Today’s tooling audits the output. But we need to be auditing decisions. How did we get there? Why did we get there?”

His whole point sits in one phrase. Compliance costume. We built tools that record what happened, then dressed them up as tools that govern what should. A security camera records the lobby. It doesn’t decide who gets in, and it can’t stop anyone. Most AI governance is the camera. The line I keep coming back to is approved is not governed.

Corruption is a systems property

Then he said the part most governance decks are built to never say out loud.

“Formal governance is designed as if everyone is acting in good faith, that the office worker will calmly fill out the right form. That’s a fantasy. It’s important to look at who can lie, who can hide the evidence, who can rewrite the record, who can offload the blame to the tool. The big mistake is treating corruption as a personality defect. In systems terms, corruption is what happens when discretion sits at the coordination layer and the evidence trail depends on human honesty.”

Read that last line twice. Discretion at the coordination layer, plus a record that runs on human honesty, gives you corruption. Not because people are bad. Because the system left the door open. Most governance starts by trusting everyone. Montgomery starts by asking who will lie, who will bury the record, who will pin it on the tool, and then he builds so they can’t. That is what separates a log from evidence. A log only works if everyone tells the truth. Evidence has to survive the person who doesn’t. Append-only signatures. Hashes. The whole decision chain. So no one can redraw the record after the fact and hand it back as the original.

Sovereignty off the cloud

I asked why almost nobody designs for air-gapped governance.

“Because air-gap governance breaks the rental model. The modern AI stack wants every serious capability to phone home, so they own your data, identity, inference, policy, logging, billing, monitoring. That makes sense if your goal is platform control. It’s catastrophic if your goal is operator sovereignty. The moment the network disappears, or the vendor’s gone, or the environment is too sensitive, like government or HIPAA, most of those guarantees collapse into nothing. The vendor gets compliance and the operator gets a black box they can’t see into.”

The people who pay for that gap can’t phone home in the first place. Hospitals. Government. Insurers. Anyone air-gapped or regulated. They get the vendor’s compliance and a black box they can’t see inside. His fix is to keep the model local and make every decision carry its own proof. He is building toward that himself, with a system that rolls back its own adaptations when they start drifting out of alignment, keyed to a hash chain. He asked me to keep the details light since it is not public yet, so I will.

One more move of his I keep chewing on. Push the guardrails all the way down to compile time.

“I write a lot of my code with AI in Rust or TypeScript because it forces the AI to actually deal with a compiler. If it doesn’t compile, it’s not working. You can still write bad code, but it forces you to at least fix what doesn’t compile. So it’s at least less broken.”

The compiler doesn’t care about good intentions. Broken code doesn’t ship because it can’t. Make the safe path the only path the model can take, and discipline stops being the thing you have to hope for.

What changed in my thinking

Two builders, opposite directions, same primitive. I build GuardSpine, and it writes hash-chained evidence bundles so anyone can verify a decision later. Montgomery, alone and self-taught, reached for a hash chain from the model side for the same reason. When two people who never compared notes build the same structure, the structure is load-bearing.

What he sharpened for me is the threat model. I started building evidence because auditors ask for proof. He builds it because people lie. His reason is the harder one, and the better one. Proof that only holds while everyone stays honest was never proof. It was a log with a nicer logo.

He also said something I did not want to hear. A lot of governance tooling will age badly as models run longer, spin up sub-agents, and learn to route around their own controls. He is right. It is exactly why I govern the decision and the artifact instead of the model. The model will not sit still. What it did, and whether that was sound, has to.

The line

“Most of it was never governance tooling. It was cloud observability with a compliance costume.”

Cloud telemetry shows that something happened. It can’t show the decision was sound. That gap is the whole job.

---

Approved is not governed. If your governance story is a dashboard, ask Montgomery’s question. Who in this room could rewrite the record after the fact, and would anyone ever know? If you want to pressure-test where your own governance is just a costume, reply or grab a slot.

David Youssef Founder, GuardSpine cal.com/davidyoussef/guardspine

---

Montgomery Kuykendall is the founder of Kuykendall Industries and a self-taught systems architect who builds AI that runs locally, off the cloud. He reviewed the piece before publication.