Introducing Crux · Cibi Sounder

Labs are automating research. OpenAI says it reached an “automated research intern” in September and wants a fully automated AI researcher by March 2028.¹ On the safety side, Anthropic, Resolution, MATS and Foresight are making the same bet for alignment research.

I think they’re right. I also think it creates a problem that isn’t solved yet.

If AI does most of the alignment research, someone has to decide which of its results are true. Today that someone is a researcher checking every step by hand. That doesn’t scale, and it gets harder every time the models get better than the people checking them.

Crux is my early attempt at this. It’s a second AI whose whole job is to prove the first one wrong.

See the deck ↗

The researcher

Julian Quick is an incoming alignment scientist at AE Studio. He already does his research with Claude Code, and he slows it down on purpose. He writes down the results he expects before a run, then checks each step against them.

He has good reason to. One agent run came back with a result that looked right but had missed something important in a paper. He only understood what went wrong after rebuilding the whole run himself. The agent saved him a fraction of the time it should have.

I asked him what he’d want if he could have anything. He said another version of himself that could do more research. He doesn’t trust models to run unsupervised yet. He wants to make the calls, not recheck every step.

What I almost built

I went into the BlueDot AI for Epistemics hackathon planning to build the research agent itself. Read the papers, run the experiments, write it up.

Then I looked at what Claude Code already does. It covers most of that, and it gets better every few months. I’d be racing the labs on their own turf, and every model release would make my project less useful.

Julian already had an agent that could do the research. He still had to check all of it by hand. So I built the checker instead.

How it works

Crux sits next to the agent a researcher already uses. The agent that produced a result never gets to sign off on it.

Check. Recount the numbers against the logs, compare the run with its plan, look for overlap between training and test data, and confirm that cited papers exist.
Attack. A second model, starting fresh and ideally from a different model family, searches the same literature with the opposite goal. It looks for missed papers, confounds and conclusions the data doesn’t support. The research agent has to fix each issue, concede it, or pass it up.
Hand off. The researcher gets the claim as it stands after review, plus at most three decisions that need their judgment.

What it catches

In the demo, an agent was asked to make a model less sycophantic. It reported back:

Sycophancy dropped from 40% to 18%. The model is now more honest.

Crux found two problems.² The model had also started disagreeing with users who were right, up from 5% to 35%. And most of the test questions were ones it had already seen in training.

The researcher would get a smaller claim they can trust and three decisions to make, instead of redoing the run.

Why I think this matters

Better models mean more AI-made research, and all of it needs checking. So a reviewer gets more useful as the labs improve their agents.

The early evidence also says automated research fails quietly. One 2026 paper found AI mistakes tend to land where human reviewers are least likely to look.³ In ResearchArena, AI monitors caught sabotage hidden in training data less than half the time.⁴ Anthropic’s automated researchers beat human baselines on an alignment task, and also gamed their own metric.⁵

I’m focused on honest mistakes, where an agent is trying its best and still reports something false.

How it could fail

Shared blind spots. If the reviewer thinks like the agent, it misses the same things. So the reviewer uses a different model and prompt, and works from the literature.
Too many nitpicks. Too many flags and the researcher stops reading. Issues will be ranked, and the researcher sees at most three.
Trusting a pass too much. Each verdict will link to its evidence. A pass means the result survived review. It doesn’t mean it’s true.

What’s next

First, Julian runs it on real research instead of my scripted demo. Then I want to pilot it with a handful of alignment researchers and measure one number: how many hours of human checking it takes to trust a result. I’ll plant known errors to see what it catches, and I’ll publish what it misses.

If you do alignment research with agents and want to try it, or you fund or advise this kind of work, email me.