The AI industry argues about alignment as if there were only two options. Either you trust the lab that built the model, or you do not use it. I think there is a third path, and it is the reason I started SAIN. I laid out the whole case in Why SAIN is building the trust infrastructure for AI. This piece is about one part of it: why the answer takes three disciplines, and why two out of three does not work.
The third path is this. Prove what the machine did without exposing what it saw.
Trusting the lab is a hope
When a firm asks whether it can put an agent on client work, the usual answer is a list of assurances. The model was tested. The vendor has policies. The data is handled carefully. All of that may be true. None of it is something the firm can check, and none of it is something the firm can hand to the next person who asks the same question.
A managing partner cannot tell a client, "We trust the vendor." A lender cannot tell an examiner that. A government contractor cannot write it in a proposal. They need something they can show.
What each discipline contributes
Intelligence is the part everyone has. An agent that reads the inbox, drafts the reply, schedules the closing, and follows up when nobody answers. Without it there is no work to account for. We build agents and ship them to paying customers, because a trust layer wrapped around a demo proves nothing.
Transparency is the record. Every action the agent takes, every approval a person gives, every draft, written to an immutable audit trail. Immutable is the point. A log that an administrator can tidy up after a bad day is a diary. A shared record that nobody can quietly edit is evidence, because every party can verify the same history.
Privacy is what makes the record usable. This is the piece people skip, and it is the piece that kills most attempts. The record of a law firm's day is made of privileged material. The record of a lender's day is made of other people's finances. Nobody responsible for that data will write it somewhere others can read it. Privacy-enhancing technology, including zero-knowledge proofs, solves exactly this. It lets you prove that something happened, or that a rule was followed, without revealing the data underneath.
Why two out of three fails
Take away privacy and you have agents and a transparent record. It works in a demo and dies in legal review, because the record itself is a disclosure.
Take away transparency and you have agents and strong privacy. Everything is locked down, and nothing can be verified by anyone outside. You are back to "trust us," with better locks.
Take away intelligence and you have a private, provable record with nothing urgent to record. I spent a decade on those two disciplines before agents existed to need them. Agents changed the stakes. For the first time, software is taking actions that carry professional liability, at volume, on behalf of people who have to answer for them.
What it looks like together
Picture a title lawyer with four inboxes. An agent works those inboxes all day. It drafts, it chases documents, it schedules. Nothing leaves without the lawyer's approval.
Six months later, a party to a closing disputes what was sent and when. The lawyer does not reconstruct the story from memory and scattered threads. The record shows each draft, each approval, and each send, in order, and it shows that nothing was altered afterward.
Now an auditor wants proof that every outgoing message was approved by a licensed professional. The firm can prove that rule was followed without handing over a single client message.
The agent did the work. The record made it defensible. Privacy made the record possible.
Where this leaves the alignment argument
I do not think the argument about what happens inside the model is going away, and I am glad serious people are working on it. But the buyers I talk to cannot wait for it to be settled, and they do not need it to be. They are not responsible for what the model thinks. They are responsible for what it did under their name.
That is a question about actions, and actions can be recorded, approved, and proven. You can hold a machine to the same standard you hold a junior colleague: show your work, get sign-off, and leave a file that survives scrutiny.
Why this combination is rare
These three fields grew up apart, with different people, different tools, and different instincts. I spent years in the first two before I built the third. That history is the reason SAIN exists, and the full argument for why it matters now is in the anchor piece.
If you work in legal, title, finance, or government contracting, the fastest way to judge this is to see it.